1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
68 #include <net/tcp_memcontrol.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
76 EXPORT_SYMBOL(memory_cgrp_subsys);
78 #define MEM_CGROUP_RECLAIM_RETRIES 5
79 static struct mem_cgroup *root_mem_cgroup __read_mostly;
80 struct cgroup_subsys_state *mem_cgroup_root_css __read_mostly;
82 /* Whether the swap controller is active */
83 #ifdef CONFIG_MEMCG_SWAP
84 int do_swap_account __read_mostly;
86 #define do_swap_account 0
89 static const char * const mem_cgroup_stat_names[] = {
99 static const char * const mem_cgroup_events_names[] = {
106 static const char * const mem_cgroup_lru_names[] = {
114 #define THRESHOLDS_EVENTS_TARGET 128
115 #define SOFTLIMIT_EVENTS_TARGET 1024
116 #define NUMAINFO_EVENTS_TARGET 1024
119 * Cgroups above their limits are maintained in a RB-Tree, independent of
120 * their hierarchy representation
123 struct mem_cgroup_tree_per_zone {
124 struct rb_root rb_root;
128 struct mem_cgroup_tree_per_node {
129 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
132 struct mem_cgroup_tree {
133 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
136 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
139 struct mem_cgroup_eventfd_list {
140 struct list_head list;
141 struct eventfd_ctx *eventfd;
145 * cgroup_event represents events which userspace want to receive.
147 struct mem_cgroup_event {
149 * memcg which the event belongs to.
151 struct mem_cgroup *memcg;
153 * eventfd to signal userspace about the event.
155 struct eventfd_ctx *eventfd;
157 * Each of these stored in a list by the cgroup.
159 struct list_head list;
161 * register_event() callback will be used to add new userspace
162 * waiter for changes related to this event. Use eventfd_signal()
163 * on eventfd to send notification to userspace.
165 int (*register_event)(struct mem_cgroup *memcg,
166 struct eventfd_ctx *eventfd, const char *args);
168 * unregister_event() callback will be called when userspace closes
169 * the eventfd or on cgroup removing. This callback must be set,
170 * if you want provide notification functionality.
172 void (*unregister_event)(struct mem_cgroup *memcg,
173 struct eventfd_ctx *eventfd);
175 * All fields below needed to unregister event when
176 * userspace closes eventfd.
179 wait_queue_head_t *wqh;
181 struct work_struct remove;
184 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
185 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
187 /* Stuffs for move charges at task migration. */
189 * Types of charges to be moved.
191 #define MOVE_ANON 0x1U
192 #define MOVE_FILE 0x2U
193 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
195 /* "mc" and its members are protected by cgroup_mutex */
196 static struct move_charge_struct {
197 spinlock_t lock; /* for from, to */
198 struct mem_cgroup *from;
199 struct mem_cgroup *to;
201 unsigned long precharge;
202 unsigned long moved_charge;
203 unsigned long moved_swap;
204 struct task_struct *moving_task; /* a task moving charges */
205 wait_queue_head_t waitq; /* a waitq for other context */
207 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
208 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
212 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
213 * limit reclaim to prevent infinite loops, if they ever occur.
215 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
216 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
219 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
220 MEM_CGROUP_CHARGE_TYPE_ANON,
221 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
222 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
226 /* for encoding cft->private value on file */
234 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
235 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
236 #define MEMFILE_ATTR(val) ((val) & 0xffff)
237 /* Used for OOM nofiier */
238 #define OOM_CONTROL (0)
241 * The memcg_create_mutex will be held whenever a new cgroup is created.
242 * As a consequence, any change that needs to protect against new child cgroups
243 * appearing has to hold it as well.
245 static DEFINE_MUTEX(memcg_create_mutex);
247 /* Some nice accessors for the vmpressure. */
248 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
251 memcg = root_mem_cgroup;
252 return &memcg->vmpressure;
255 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
257 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
260 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
262 return (memcg == root_mem_cgroup);
266 * We restrict the id in the range of [1, 65535], so it can fit into
269 #define MEM_CGROUP_ID_MAX USHRT_MAX
271 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
273 return memcg->css.id;
277 * A helper function to get mem_cgroup from ID. must be called under
278 * rcu_read_lock(). The caller is responsible for calling
279 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
280 * refcnt from swap can be called against removed memcg.)
282 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
284 struct cgroup_subsys_state *css;
286 css = css_from_id(id, &memory_cgrp_subsys);
287 return mem_cgroup_from_css(css);
290 /* Writing them here to avoid exposing memcg's inner layout */
291 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
293 void sock_update_memcg(struct sock *sk)
295 if (mem_cgroup_sockets_enabled) {
296 struct mem_cgroup *memcg;
297 struct cg_proto *cg_proto;
299 BUG_ON(!sk->sk_prot->proto_cgroup);
301 /* Socket cloning can throw us here with sk_cgrp already
302 * filled. It won't however, necessarily happen from
303 * process context. So the test for root memcg given
304 * the current task's memcg won't help us in this case.
306 * Respecting the original socket's memcg is a better
307 * decision in this case.
310 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
311 css_get(&sk->sk_cgrp->memcg->css);
316 memcg = mem_cgroup_from_task(current);
317 cg_proto = sk->sk_prot->proto_cgroup(memcg);
318 if (cg_proto && test_bit(MEMCG_SOCK_ACTIVE, &cg_proto->flags) &&
319 css_tryget_online(&memcg->css)) {
320 sk->sk_cgrp = cg_proto;
325 EXPORT_SYMBOL(sock_update_memcg);
327 void sock_release_memcg(struct sock *sk)
329 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
330 struct mem_cgroup *memcg;
331 WARN_ON(!sk->sk_cgrp->memcg);
332 memcg = sk->sk_cgrp->memcg;
333 css_put(&sk->sk_cgrp->memcg->css);
337 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
339 if (!memcg || mem_cgroup_is_root(memcg))
342 return &memcg->tcp_mem;
344 EXPORT_SYMBOL(tcp_proto_cgroup);
348 #ifdef CONFIG_MEMCG_KMEM
350 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
351 * The main reason for not using cgroup id for this:
352 * this works better in sparse environments, where we have a lot of memcgs,
353 * but only a few kmem-limited. Or also, if we have, for instance, 200
354 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
355 * 200 entry array for that.
357 * The current size of the caches array is stored in memcg_nr_cache_ids. It
358 * will double each time we have to increase it.
360 static DEFINE_IDA(memcg_cache_ida);
361 int memcg_nr_cache_ids;
363 /* Protects memcg_nr_cache_ids */
364 static DECLARE_RWSEM(memcg_cache_ids_sem);
366 void memcg_get_cache_ids(void)
368 down_read(&memcg_cache_ids_sem);
371 void memcg_put_cache_ids(void)
373 up_read(&memcg_cache_ids_sem);
377 * MIN_SIZE is different than 1, because we would like to avoid going through
378 * the alloc/free process all the time. In a small machine, 4 kmem-limited
379 * cgroups is a reasonable guess. In the future, it could be a parameter or
380 * tunable, but that is strictly not necessary.
382 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
383 * this constant directly from cgroup, but it is understandable that this is
384 * better kept as an internal representation in cgroup.c. In any case, the
385 * cgrp_id space is not getting any smaller, and we don't have to necessarily
386 * increase ours as well if it increases.
388 #define MEMCG_CACHES_MIN_SIZE 4
389 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
392 * A lot of the calls to the cache allocation functions are expected to be
393 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
394 * conditional to this static branch, we'll have to allow modules that does
395 * kmem_cache_alloc and the such to see this symbol as well
397 struct static_key memcg_kmem_enabled_key;
398 EXPORT_SYMBOL(memcg_kmem_enabled_key);
400 #endif /* CONFIG_MEMCG_KMEM */
402 static struct mem_cgroup_per_zone *
403 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
405 int nid = zone_to_nid(zone);
406 int zid = zone_idx(zone);
408 return &memcg->nodeinfo[nid]->zoneinfo[zid];
412 * mem_cgroup_css_from_page - css of the memcg associated with a page
413 * @page: page of interest
415 * If memcg is bound to the default hierarchy, css of the memcg associated
416 * with @page is returned. The returned css remains associated with @page
417 * until it is released.
419 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
422 * XXX: The above description of behavior on the default hierarchy isn't
423 * strictly true yet as replace_page_cache_page() can modify the
424 * association before @page is released even on the default hierarchy;
425 * however, the current and planned usages don't mix the the two functions
426 * and replace_page_cache_page() will soon be updated to make the invariant
429 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
431 struct mem_cgroup *memcg;
435 memcg = page->mem_cgroup;
437 if (!memcg || !cgroup_on_dfl(memcg->css.cgroup))
438 memcg = root_mem_cgroup;
444 static struct mem_cgroup_per_zone *
445 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
447 int nid = page_to_nid(page);
448 int zid = page_zonenum(page);
450 return &memcg->nodeinfo[nid]->zoneinfo[zid];
453 static struct mem_cgroup_tree_per_zone *
454 soft_limit_tree_node_zone(int nid, int zid)
456 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
459 static struct mem_cgroup_tree_per_zone *
460 soft_limit_tree_from_page(struct page *page)
462 int nid = page_to_nid(page);
463 int zid = page_zonenum(page);
465 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
468 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
469 struct mem_cgroup_tree_per_zone *mctz,
470 unsigned long new_usage_in_excess)
472 struct rb_node **p = &mctz->rb_root.rb_node;
473 struct rb_node *parent = NULL;
474 struct mem_cgroup_per_zone *mz_node;
479 mz->usage_in_excess = new_usage_in_excess;
480 if (!mz->usage_in_excess)
484 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
486 if (mz->usage_in_excess < mz_node->usage_in_excess)
489 * We can't avoid mem cgroups that are over their soft
490 * limit by the same amount
492 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
495 rb_link_node(&mz->tree_node, parent, p);
496 rb_insert_color(&mz->tree_node, &mctz->rb_root);
500 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
501 struct mem_cgroup_tree_per_zone *mctz)
505 rb_erase(&mz->tree_node, &mctz->rb_root);
509 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
510 struct mem_cgroup_tree_per_zone *mctz)
514 spin_lock_irqsave(&mctz->lock, flags);
515 __mem_cgroup_remove_exceeded(mz, mctz);
516 spin_unlock_irqrestore(&mctz->lock, flags);
519 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
521 unsigned long nr_pages = page_counter_read(&memcg->memory);
522 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
523 unsigned long excess = 0;
525 if (nr_pages > soft_limit)
526 excess = nr_pages - soft_limit;
531 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
533 unsigned long excess;
534 struct mem_cgroup_per_zone *mz;
535 struct mem_cgroup_tree_per_zone *mctz;
537 mctz = soft_limit_tree_from_page(page);
539 * Necessary to update all ancestors when hierarchy is used.
540 * because their event counter is not touched.
542 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
543 mz = mem_cgroup_page_zoneinfo(memcg, page);
544 excess = soft_limit_excess(memcg);
546 * We have to update the tree if mz is on RB-tree or
547 * mem is over its softlimit.
549 if (excess || mz->on_tree) {
552 spin_lock_irqsave(&mctz->lock, flags);
553 /* if on-tree, remove it */
555 __mem_cgroup_remove_exceeded(mz, mctz);
557 * Insert again. mz->usage_in_excess will be updated.
558 * If excess is 0, no tree ops.
560 __mem_cgroup_insert_exceeded(mz, mctz, excess);
561 spin_unlock_irqrestore(&mctz->lock, flags);
566 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
568 struct mem_cgroup_tree_per_zone *mctz;
569 struct mem_cgroup_per_zone *mz;
573 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
574 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
575 mctz = soft_limit_tree_node_zone(nid, zid);
576 mem_cgroup_remove_exceeded(mz, mctz);
581 static struct mem_cgroup_per_zone *
582 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
584 struct rb_node *rightmost = NULL;
585 struct mem_cgroup_per_zone *mz;
589 rightmost = rb_last(&mctz->rb_root);
591 goto done; /* Nothing to reclaim from */
593 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
595 * Remove the node now but someone else can add it back,
596 * we will to add it back at the end of reclaim to its correct
597 * position in the tree.
599 __mem_cgroup_remove_exceeded(mz, mctz);
600 if (!soft_limit_excess(mz->memcg) ||
601 !css_tryget_online(&mz->memcg->css))
607 static struct mem_cgroup_per_zone *
608 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
610 struct mem_cgroup_per_zone *mz;
612 spin_lock_irq(&mctz->lock);
613 mz = __mem_cgroup_largest_soft_limit_node(mctz);
614 spin_unlock_irq(&mctz->lock);
619 * Implementation Note: reading percpu statistics for memcg.
621 * Both of vmstat[] and percpu_counter has threshold and do periodic
622 * synchronization to implement "quick" read. There are trade-off between
623 * reading cost and precision of value. Then, we may have a chance to implement
624 * a periodic synchronizion of counter in memcg's counter.
626 * But this _read() function is used for user interface now. The user accounts
627 * memory usage by memory cgroup and he _always_ requires exact value because
628 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
629 * have to visit all online cpus and make sum. So, for now, unnecessary
630 * synchronization is not implemented. (just implemented for cpu hotplug)
632 * If there are kernel internal actions which can make use of some not-exact
633 * value, and reading all cpu value can be performance bottleneck in some
634 * common workload, threashold and synchonization as vmstat[] should be
637 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
638 enum mem_cgroup_stat_index idx)
643 for_each_possible_cpu(cpu)
644 val += per_cpu(memcg->stat->count[idx], cpu);
648 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
649 enum mem_cgroup_events_index idx)
651 unsigned long val = 0;
654 for_each_possible_cpu(cpu)
655 val += per_cpu(memcg->stat->events[idx], cpu);
659 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
664 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
665 * counted as CACHE even if it's on ANON LRU.
668 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
671 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
674 if (PageTransHuge(page))
675 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
678 /* pagein of a big page is an event. So, ignore page size */
680 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
682 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
683 nr_pages = -nr_pages; /* for event */
686 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
689 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
691 unsigned int lru_mask)
693 unsigned long nr = 0;
696 VM_BUG_ON((unsigned)nid >= nr_node_ids);
698 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
699 struct mem_cgroup_per_zone *mz;
703 if (!(BIT(lru) & lru_mask))
705 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
706 nr += mz->lru_size[lru];
712 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
713 unsigned int lru_mask)
715 unsigned long nr = 0;
718 for_each_node_state(nid, N_MEMORY)
719 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
723 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
724 enum mem_cgroup_events_target target)
726 unsigned long val, next;
728 val = __this_cpu_read(memcg->stat->nr_page_events);
729 next = __this_cpu_read(memcg->stat->targets[target]);
730 /* from time_after() in jiffies.h */
731 if ((long)next - (long)val < 0) {
733 case MEM_CGROUP_TARGET_THRESH:
734 next = val + THRESHOLDS_EVENTS_TARGET;
736 case MEM_CGROUP_TARGET_SOFTLIMIT:
737 next = val + SOFTLIMIT_EVENTS_TARGET;
739 case MEM_CGROUP_TARGET_NUMAINFO:
740 next = val + NUMAINFO_EVENTS_TARGET;
745 __this_cpu_write(memcg->stat->targets[target], next);
752 * Check events in order.
755 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
757 /* threshold event is triggered in finer grain than soft limit */
758 if (unlikely(mem_cgroup_event_ratelimit(memcg,
759 MEM_CGROUP_TARGET_THRESH))) {
761 bool do_numainfo __maybe_unused;
763 do_softlimit = mem_cgroup_event_ratelimit(memcg,
764 MEM_CGROUP_TARGET_SOFTLIMIT);
766 do_numainfo = mem_cgroup_event_ratelimit(memcg,
767 MEM_CGROUP_TARGET_NUMAINFO);
769 mem_cgroup_threshold(memcg);
770 if (unlikely(do_softlimit))
771 mem_cgroup_update_tree(memcg, page);
773 if (unlikely(do_numainfo))
774 atomic_inc(&memcg->numainfo_events);
779 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
782 * mm_update_next_owner() may clear mm->owner to NULL
783 * if it races with swapoff, page migration, etc.
784 * So this can be called with p == NULL.
789 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
791 EXPORT_SYMBOL(mem_cgroup_from_task);
793 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
795 struct mem_cgroup *memcg = NULL;
800 * Page cache insertions can happen withou an
801 * actual mm context, e.g. during disk probing
802 * on boot, loopback IO, acct() writes etc.
805 memcg = root_mem_cgroup;
807 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
808 if (unlikely(!memcg))
809 memcg = root_mem_cgroup;
811 } while (!css_tryget_online(&memcg->css));
817 * mem_cgroup_iter - iterate over memory cgroup hierarchy
818 * @root: hierarchy root
819 * @prev: previously returned memcg, NULL on first invocation
820 * @reclaim: cookie for shared reclaim walks, NULL for full walks
822 * Returns references to children of the hierarchy below @root, or
823 * @root itself, or %NULL after a full round-trip.
825 * Caller must pass the return value in @prev on subsequent
826 * invocations for reference counting, or use mem_cgroup_iter_break()
827 * to cancel a hierarchy walk before the round-trip is complete.
829 * Reclaimers can specify a zone and a priority level in @reclaim to
830 * divide up the memcgs in the hierarchy among all concurrent
831 * reclaimers operating on the same zone and priority.
833 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
834 struct mem_cgroup *prev,
835 struct mem_cgroup_reclaim_cookie *reclaim)
837 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
838 struct cgroup_subsys_state *css = NULL;
839 struct mem_cgroup *memcg = NULL;
840 struct mem_cgroup *pos = NULL;
842 if (mem_cgroup_disabled())
846 root = root_mem_cgroup;
848 if (prev && !reclaim)
851 if (!root->use_hierarchy && root != root_mem_cgroup) {
860 struct mem_cgroup_per_zone *mz;
862 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
863 iter = &mz->iter[reclaim->priority];
865 if (prev && reclaim->generation != iter->generation)
869 pos = READ_ONCE(iter->position);
871 * A racing update may change the position and
872 * put the last reference, hence css_tryget(),
873 * or retry to see the updated position.
875 } while (pos && !css_tryget(&pos->css));
882 css = css_next_descendant_pre(css, &root->css);
885 * Reclaimers share the hierarchy walk, and a
886 * new one might jump in right at the end of
887 * the hierarchy - make sure they see at least
888 * one group and restart from the beginning.
896 * Verify the css and acquire a reference. The root
897 * is provided by the caller, so we know it's alive
898 * and kicking, and don't take an extra reference.
900 memcg = mem_cgroup_from_css(css);
902 if (css == &root->css)
905 if (css_tryget(css)) {
907 * Make sure the memcg is initialized:
908 * mem_cgroup_css_online() orders the the
909 * initialization against setting the flag.
911 if (smp_load_acquire(&memcg->initialized))
921 if (cmpxchg(&iter->position, pos, memcg) == pos) {
923 css_get(&memcg->css);
929 * pairs with css_tryget when dereferencing iter->position
938 reclaim->generation = iter->generation;
944 if (prev && prev != root)
951 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
952 * @root: hierarchy root
953 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
955 void mem_cgroup_iter_break(struct mem_cgroup *root,
956 struct mem_cgroup *prev)
959 root = root_mem_cgroup;
960 if (prev && prev != root)
965 * Iteration constructs for visiting all cgroups (under a tree). If
966 * loops are exited prematurely (break), mem_cgroup_iter_break() must
967 * be used for reference counting.
969 #define for_each_mem_cgroup_tree(iter, root) \
970 for (iter = mem_cgroup_iter(root, NULL, NULL); \
972 iter = mem_cgroup_iter(root, iter, NULL))
974 #define for_each_mem_cgroup(iter) \
975 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
977 iter = mem_cgroup_iter(NULL, iter, NULL))
980 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
981 * @zone: zone of the wanted lruvec
982 * @memcg: memcg of the wanted lruvec
984 * Returns the lru list vector holding pages for the given @zone and
985 * @mem. This can be the global zone lruvec, if the memory controller
988 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
989 struct mem_cgroup *memcg)
991 struct mem_cgroup_per_zone *mz;
992 struct lruvec *lruvec;
994 if (mem_cgroup_disabled()) {
995 lruvec = &zone->lruvec;
999 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1000 lruvec = &mz->lruvec;
1003 * Since a node can be onlined after the mem_cgroup was created,
1004 * we have to be prepared to initialize lruvec->zone here;
1005 * and if offlined then reonlined, we need to reinitialize it.
1007 if (unlikely(lruvec->zone != zone))
1008 lruvec->zone = zone;
1013 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1015 * @zone: zone of the page
1017 * This function is only safe when following the LRU page isolation
1018 * and putback protocol: the LRU lock must be held, and the page must
1019 * either be PageLRU() or the caller must have isolated/allocated it.
1021 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1023 struct mem_cgroup_per_zone *mz;
1024 struct mem_cgroup *memcg;
1025 struct lruvec *lruvec;
1027 if (mem_cgroup_disabled()) {
1028 lruvec = &zone->lruvec;
1032 memcg = page->mem_cgroup;
1034 * Swapcache readahead pages are added to the LRU - and
1035 * possibly migrated - before they are charged.
1038 memcg = root_mem_cgroup;
1040 mz = mem_cgroup_page_zoneinfo(memcg, page);
1041 lruvec = &mz->lruvec;
1044 * Since a node can be onlined after the mem_cgroup was created,
1045 * we have to be prepared to initialize lruvec->zone here;
1046 * and if offlined then reonlined, we need to reinitialize it.
1048 if (unlikely(lruvec->zone != zone))
1049 lruvec->zone = zone;
1054 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1055 * @lruvec: mem_cgroup per zone lru vector
1056 * @lru: index of lru list the page is sitting on
1057 * @nr_pages: positive when adding or negative when removing
1059 * This function must be called when a page is added to or removed from an
1062 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1065 struct mem_cgroup_per_zone *mz;
1066 unsigned long *lru_size;
1068 if (mem_cgroup_disabled())
1071 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1072 lru_size = mz->lru_size + lru;
1073 *lru_size += nr_pages;
1074 VM_BUG_ON((long)(*lru_size) < 0);
1077 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1079 struct mem_cgroup *task_memcg;
1080 struct task_struct *p;
1083 p = find_lock_task_mm(task);
1085 task_memcg = get_mem_cgroup_from_mm(p->mm);
1089 * All threads may have already detached their mm's, but the oom
1090 * killer still needs to detect if they have already been oom
1091 * killed to prevent needlessly killing additional tasks.
1094 task_memcg = mem_cgroup_from_task(task);
1095 css_get(&task_memcg->css);
1098 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1099 css_put(&task_memcg->css);
1103 #define mem_cgroup_from_counter(counter, member) \
1104 container_of(counter, struct mem_cgroup, member)
1107 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1108 * @memcg: the memory cgroup
1110 * Returns the maximum amount of memory @mem can be charged with, in
1113 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1115 unsigned long margin = 0;
1116 unsigned long count;
1117 unsigned long limit;
1119 count = page_counter_read(&memcg->memory);
1120 limit = READ_ONCE(memcg->memory.limit);
1122 margin = limit - count;
1124 if (do_swap_account) {
1125 count = page_counter_read(&memcg->memsw);
1126 limit = READ_ONCE(memcg->memsw.limit);
1128 margin = min(margin, limit - count);
1135 * A routine for checking "mem" is under move_account() or not.
1137 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1138 * moving cgroups. This is for waiting at high-memory pressure
1141 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1143 struct mem_cgroup *from;
1144 struct mem_cgroup *to;
1147 * Unlike task_move routines, we access mc.to, mc.from not under
1148 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1150 spin_lock(&mc.lock);
1156 ret = mem_cgroup_is_descendant(from, memcg) ||
1157 mem_cgroup_is_descendant(to, memcg);
1159 spin_unlock(&mc.lock);
1163 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1165 if (mc.moving_task && current != mc.moving_task) {
1166 if (mem_cgroup_under_move(memcg)) {
1168 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1169 /* moving charge context might have finished. */
1172 finish_wait(&mc.waitq, &wait);
1179 #define K(x) ((x) << (PAGE_SHIFT-10))
1181 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1182 * @memcg: The memory cgroup that went over limit
1183 * @p: Task that is going to be killed
1185 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1188 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1190 /* oom_info_lock ensures that parallel ooms do not interleave */
1191 static DEFINE_MUTEX(oom_info_lock);
1192 struct mem_cgroup *iter;
1195 mutex_lock(&oom_info_lock);
1199 pr_info("Task in ");
1200 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1201 pr_cont(" killed as a result of limit of ");
1203 pr_info("Memory limit reached of cgroup ");
1206 pr_cont_cgroup_path(memcg->css.cgroup);
1211 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1212 K((u64)page_counter_read(&memcg->memory)),
1213 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1214 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1215 K((u64)page_counter_read(&memcg->memsw)),
1216 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1217 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1218 K((u64)page_counter_read(&memcg->kmem)),
1219 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1221 for_each_mem_cgroup_tree(iter, memcg) {
1222 pr_info("Memory cgroup stats for ");
1223 pr_cont_cgroup_path(iter->css.cgroup);
1226 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1227 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1229 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1230 K(mem_cgroup_read_stat(iter, i)));
1233 for (i = 0; i < NR_LRU_LISTS; i++)
1234 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1235 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1239 mutex_unlock(&oom_info_lock);
1243 * This function returns the number of memcg under hierarchy tree. Returns
1244 * 1(self count) if no children.
1246 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1249 struct mem_cgroup *iter;
1251 for_each_mem_cgroup_tree(iter, memcg)
1257 * Return the memory (and swap, if configured) limit for a memcg.
1259 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1261 unsigned long limit;
1263 limit = memcg->memory.limit;
1264 if (mem_cgroup_swappiness(memcg)) {
1265 unsigned long memsw_limit;
1267 memsw_limit = memcg->memsw.limit;
1268 limit = min(limit + total_swap_pages, memsw_limit);
1273 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1276 struct oom_control oc = {
1279 .gfp_mask = gfp_mask,
1282 struct mem_cgroup *iter;
1283 unsigned long chosen_points = 0;
1284 unsigned long totalpages;
1285 unsigned int points = 0;
1286 struct task_struct *chosen = NULL;
1288 mutex_lock(&oom_lock);
1291 * If current has a pending SIGKILL or is exiting, then automatically
1292 * select it. The goal is to allow it to allocate so that it may
1293 * quickly exit and free its memory.
1295 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1296 mark_oom_victim(current);
1300 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1301 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1302 for_each_mem_cgroup_tree(iter, memcg) {
1303 struct css_task_iter it;
1304 struct task_struct *task;
1306 css_task_iter_start(&iter->css, &it);
1307 while ((task = css_task_iter_next(&it))) {
1308 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1309 case OOM_SCAN_SELECT:
1311 put_task_struct(chosen);
1313 chosen_points = ULONG_MAX;
1314 get_task_struct(chosen);
1316 case OOM_SCAN_CONTINUE:
1318 case OOM_SCAN_ABORT:
1319 css_task_iter_end(&it);
1320 mem_cgroup_iter_break(memcg, iter);
1322 put_task_struct(chosen);
1327 points = oom_badness(task, memcg, NULL, totalpages);
1328 if (!points || points < chosen_points)
1330 /* Prefer thread group leaders for display purposes */
1331 if (points == chosen_points &&
1332 thread_group_leader(chosen))
1336 put_task_struct(chosen);
1338 chosen_points = points;
1339 get_task_struct(chosen);
1341 css_task_iter_end(&it);
1345 points = chosen_points * 1000 / totalpages;
1346 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1347 "Memory cgroup out of memory");
1350 mutex_unlock(&oom_lock);
1353 #if MAX_NUMNODES > 1
1356 * test_mem_cgroup_node_reclaimable
1357 * @memcg: the target memcg
1358 * @nid: the node ID to be checked.
1359 * @noswap : specify true here if the user wants flle only information.
1361 * This function returns whether the specified memcg contains any
1362 * reclaimable pages on a node. Returns true if there are any reclaimable
1363 * pages in the node.
1365 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1366 int nid, bool noswap)
1368 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1370 if (noswap || !total_swap_pages)
1372 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1379 * Always updating the nodemask is not very good - even if we have an empty
1380 * list or the wrong list here, we can start from some node and traverse all
1381 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1384 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1388 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1389 * pagein/pageout changes since the last update.
1391 if (!atomic_read(&memcg->numainfo_events))
1393 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1396 /* make a nodemask where this memcg uses memory from */
1397 memcg->scan_nodes = node_states[N_MEMORY];
1399 for_each_node_mask(nid, node_states[N_MEMORY]) {
1401 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1402 node_clear(nid, memcg->scan_nodes);
1405 atomic_set(&memcg->numainfo_events, 0);
1406 atomic_set(&memcg->numainfo_updating, 0);
1410 * Selecting a node where we start reclaim from. Because what we need is just
1411 * reducing usage counter, start from anywhere is O,K. Considering
1412 * memory reclaim from current node, there are pros. and cons.
1414 * Freeing memory from current node means freeing memory from a node which
1415 * we'll use or we've used. So, it may make LRU bad. And if several threads
1416 * hit limits, it will see a contention on a node. But freeing from remote
1417 * node means more costs for memory reclaim because of memory latency.
1419 * Now, we use round-robin. Better algorithm is welcomed.
1421 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1425 mem_cgroup_may_update_nodemask(memcg);
1426 node = memcg->last_scanned_node;
1428 node = next_node(node, memcg->scan_nodes);
1429 if (node == MAX_NUMNODES)
1430 node = first_node(memcg->scan_nodes);
1432 * We call this when we hit limit, not when pages are added to LRU.
1433 * No LRU may hold pages because all pages are UNEVICTABLE or
1434 * memcg is too small and all pages are not on LRU. In that case,
1435 * we use curret node.
1437 if (unlikely(node == MAX_NUMNODES))
1438 node = numa_node_id();
1440 memcg->last_scanned_node = node;
1444 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1450 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1453 unsigned long *total_scanned)
1455 struct mem_cgroup *victim = NULL;
1458 unsigned long excess;
1459 unsigned long nr_scanned;
1460 struct mem_cgroup_reclaim_cookie reclaim = {
1465 excess = soft_limit_excess(root_memcg);
1468 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1473 * If we have not been able to reclaim
1474 * anything, it might because there are
1475 * no reclaimable pages under this hierarchy
1480 * We want to do more targeted reclaim.
1481 * excess >> 2 is not to excessive so as to
1482 * reclaim too much, nor too less that we keep
1483 * coming back to reclaim from this cgroup
1485 if (total >= (excess >> 2) ||
1486 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1491 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1493 *total_scanned += nr_scanned;
1494 if (!soft_limit_excess(root_memcg))
1497 mem_cgroup_iter_break(root_memcg, victim);
1501 #ifdef CONFIG_LOCKDEP
1502 static struct lockdep_map memcg_oom_lock_dep_map = {
1503 .name = "memcg_oom_lock",
1507 static DEFINE_SPINLOCK(memcg_oom_lock);
1510 * Check OOM-Killer is already running under our hierarchy.
1511 * If someone is running, return false.
1513 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1515 struct mem_cgroup *iter, *failed = NULL;
1517 spin_lock(&memcg_oom_lock);
1519 for_each_mem_cgroup_tree(iter, memcg) {
1520 if (iter->oom_lock) {
1522 * this subtree of our hierarchy is already locked
1523 * so we cannot give a lock.
1526 mem_cgroup_iter_break(memcg, iter);
1529 iter->oom_lock = true;
1534 * OK, we failed to lock the whole subtree so we have
1535 * to clean up what we set up to the failing subtree
1537 for_each_mem_cgroup_tree(iter, memcg) {
1538 if (iter == failed) {
1539 mem_cgroup_iter_break(memcg, iter);
1542 iter->oom_lock = false;
1545 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1547 spin_unlock(&memcg_oom_lock);
1552 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1554 struct mem_cgroup *iter;
1556 spin_lock(&memcg_oom_lock);
1557 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1558 for_each_mem_cgroup_tree(iter, memcg)
1559 iter->oom_lock = false;
1560 spin_unlock(&memcg_oom_lock);
1563 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1565 struct mem_cgroup *iter;
1567 spin_lock(&memcg_oom_lock);
1568 for_each_mem_cgroup_tree(iter, memcg)
1570 spin_unlock(&memcg_oom_lock);
1573 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1575 struct mem_cgroup *iter;
1578 * When a new child is created while the hierarchy is under oom,
1579 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1581 spin_lock(&memcg_oom_lock);
1582 for_each_mem_cgroup_tree(iter, memcg)
1583 if (iter->under_oom > 0)
1585 spin_unlock(&memcg_oom_lock);
1588 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1590 struct oom_wait_info {
1591 struct mem_cgroup *memcg;
1595 static int memcg_oom_wake_function(wait_queue_t *wait,
1596 unsigned mode, int sync, void *arg)
1598 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1599 struct mem_cgroup *oom_wait_memcg;
1600 struct oom_wait_info *oom_wait_info;
1602 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1603 oom_wait_memcg = oom_wait_info->memcg;
1605 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1606 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1608 return autoremove_wake_function(wait, mode, sync, arg);
1611 static void memcg_oom_recover(struct mem_cgroup *memcg)
1614 * For the following lockless ->under_oom test, the only required
1615 * guarantee is that it must see the state asserted by an OOM when
1616 * this function is called as a result of userland actions
1617 * triggered by the notification of the OOM. This is trivially
1618 * achieved by invoking mem_cgroup_mark_under_oom() before
1619 * triggering notification.
1621 if (memcg && memcg->under_oom)
1622 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1625 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1627 if (!current->memcg_oom.may_oom)
1630 * We are in the middle of the charge context here, so we
1631 * don't want to block when potentially sitting on a callstack
1632 * that holds all kinds of filesystem and mm locks.
1634 * Also, the caller may handle a failed allocation gracefully
1635 * (like optional page cache readahead) and so an OOM killer
1636 * invocation might not even be necessary.
1638 * That's why we don't do anything here except remember the
1639 * OOM context and then deal with it at the end of the page
1640 * fault when the stack is unwound, the locks are released,
1641 * and when we know whether the fault was overall successful.
1643 css_get(&memcg->css);
1644 current->memcg_oom.memcg = memcg;
1645 current->memcg_oom.gfp_mask = mask;
1646 current->memcg_oom.order = order;
1650 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1651 * @handle: actually kill/wait or just clean up the OOM state
1653 * This has to be called at the end of a page fault if the memcg OOM
1654 * handler was enabled.
1656 * Memcg supports userspace OOM handling where failed allocations must
1657 * sleep on a waitqueue until the userspace task resolves the
1658 * situation. Sleeping directly in the charge context with all kinds
1659 * of locks held is not a good idea, instead we remember an OOM state
1660 * in the task and mem_cgroup_oom_synchronize() has to be called at
1661 * the end of the page fault to complete the OOM handling.
1663 * Returns %true if an ongoing memcg OOM situation was detected and
1664 * completed, %false otherwise.
1666 bool mem_cgroup_oom_synchronize(bool handle)
1668 struct mem_cgroup *memcg = current->memcg_oom.memcg;
1669 struct oom_wait_info owait;
1672 /* OOM is global, do not handle */
1676 if (!handle || oom_killer_disabled)
1679 owait.memcg = memcg;
1680 owait.wait.flags = 0;
1681 owait.wait.func = memcg_oom_wake_function;
1682 owait.wait.private = current;
1683 INIT_LIST_HEAD(&owait.wait.task_list);
1685 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1686 mem_cgroup_mark_under_oom(memcg);
1688 locked = mem_cgroup_oom_trylock(memcg);
1691 mem_cgroup_oom_notify(memcg);
1693 if (locked && !memcg->oom_kill_disable) {
1694 mem_cgroup_unmark_under_oom(memcg);
1695 finish_wait(&memcg_oom_waitq, &owait.wait);
1696 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
1697 current->memcg_oom.order);
1700 mem_cgroup_unmark_under_oom(memcg);
1701 finish_wait(&memcg_oom_waitq, &owait.wait);
1705 mem_cgroup_oom_unlock(memcg);
1707 * There is no guarantee that an OOM-lock contender
1708 * sees the wakeups triggered by the OOM kill
1709 * uncharges. Wake any sleepers explicitely.
1711 memcg_oom_recover(memcg);
1714 current->memcg_oom.memcg = NULL;
1715 css_put(&memcg->css);
1720 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1721 * @page: page that is going to change accounted state
1723 * This function must mark the beginning of an accounted page state
1724 * change to prevent double accounting when the page is concurrently
1725 * being moved to another memcg:
1727 * memcg = mem_cgroup_begin_page_stat(page);
1728 * if (TestClearPageState(page))
1729 * mem_cgroup_update_page_stat(memcg, state, -1);
1730 * mem_cgroup_end_page_stat(memcg);
1732 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1734 struct mem_cgroup *memcg;
1735 unsigned long flags;
1738 * The RCU lock is held throughout the transaction. The fast
1739 * path can get away without acquiring the memcg->move_lock
1740 * because page moving starts with an RCU grace period.
1742 * The RCU lock also protects the memcg from being freed when
1743 * the page state that is going to change is the only thing
1744 * preventing the page from being uncharged.
1745 * E.g. end-writeback clearing PageWriteback(), which allows
1746 * migration to go ahead and uncharge the page before the
1747 * account transaction might be complete.
1751 if (mem_cgroup_disabled())
1754 memcg = page->mem_cgroup;
1755 if (unlikely(!memcg))
1758 if (atomic_read(&memcg->moving_account) <= 0)
1761 spin_lock_irqsave(&memcg->move_lock, flags);
1762 if (memcg != page->mem_cgroup) {
1763 spin_unlock_irqrestore(&memcg->move_lock, flags);
1768 * When charge migration first begins, we can have locked and
1769 * unlocked page stat updates happening concurrently. Track
1770 * the task who has the lock for mem_cgroup_end_page_stat().
1772 memcg->move_lock_task = current;
1773 memcg->move_lock_flags = flags;
1777 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
1780 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1781 * @memcg: the memcg that was accounted against
1783 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
1785 if (memcg && memcg->move_lock_task == current) {
1786 unsigned long flags = memcg->move_lock_flags;
1788 memcg->move_lock_task = NULL;
1789 memcg->move_lock_flags = 0;
1791 spin_unlock_irqrestore(&memcg->move_lock, flags);
1796 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
1799 * size of first charge trial. "32" comes from vmscan.c's magic value.
1800 * TODO: maybe necessary to use big numbers in big irons.
1802 #define CHARGE_BATCH 32U
1803 struct memcg_stock_pcp {
1804 struct mem_cgroup *cached; /* this never be root cgroup */
1805 unsigned int nr_pages;
1806 struct work_struct work;
1807 unsigned long flags;
1808 #define FLUSHING_CACHED_CHARGE 0
1810 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1811 static DEFINE_MUTEX(percpu_charge_mutex);
1814 * consume_stock: Try to consume stocked charge on this cpu.
1815 * @memcg: memcg to consume from.
1816 * @nr_pages: how many pages to charge.
1818 * The charges will only happen if @memcg matches the current cpu's memcg
1819 * stock, and at least @nr_pages are available in that stock. Failure to
1820 * service an allocation will refill the stock.
1822 * returns true if successful, false otherwise.
1824 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1826 struct memcg_stock_pcp *stock;
1829 if (nr_pages > CHARGE_BATCH)
1832 stock = &get_cpu_var(memcg_stock);
1833 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1834 stock->nr_pages -= nr_pages;
1837 put_cpu_var(memcg_stock);
1842 * Returns stocks cached in percpu and reset cached information.
1844 static void drain_stock(struct memcg_stock_pcp *stock)
1846 struct mem_cgroup *old = stock->cached;
1848 if (stock->nr_pages) {
1849 page_counter_uncharge(&old->memory, stock->nr_pages);
1850 if (do_swap_account)
1851 page_counter_uncharge(&old->memsw, stock->nr_pages);
1852 css_put_many(&old->css, stock->nr_pages);
1853 stock->nr_pages = 0;
1855 stock->cached = NULL;
1859 * This must be called under preempt disabled or must be called by
1860 * a thread which is pinned to local cpu.
1862 static void drain_local_stock(struct work_struct *dummy)
1864 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1866 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1870 * Cache charges(val) to local per_cpu area.
1871 * This will be consumed by consume_stock() function, later.
1873 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1875 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1877 if (stock->cached != memcg) { /* reset if necessary */
1879 stock->cached = memcg;
1881 stock->nr_pages += nr_pages;
1882 put_cpu_var(memcg_stock);
1886 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1887 * of the hierarchy under it.
1889 static void drain_all_stock(struct mem_cgroup *root_memcg)
1893 /* If someone's already draining, avoid adding running more workers. */
1894 if (!mutex_trylock(&percpu_charge_mutex))
1896 /* Notify other cpus that system-wide "drain" is running */
1899 for_each_online_cpu(cpu) {
1900 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1901 struct mem_cgroup *memcg;
1903 memcg = stock->cached;
1904 if (!memcg || !stock->nr_pages)
1906 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1908 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1910 drain_local_stock(&stock->work);
1912 schedule_work_on(cpu, &stock->work);
1917 mutex_unlock(&percpu_charge_mutex);
1920 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1921 unsigned long action,
1924 int cpu = (unsigned long)hcpu;
1925 struct memcg_stock_pcp *stock;
1927 if (action == CPU_ONLINE)
1930 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1933 stock = &per_cpu(memcg_stock, cpu);
1938 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1939 unsigned int nr_pages)
1941 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1942 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1943 struct mem_cgroup *mem_over_limit;
1944 struct page_counter *counter;
1945 unsigned long nr_reclaimed;
1946 bool may_swap = true;
1947 bool drained = false;
1950 if (mem_cgroup_is_root(memcg))
1953 if (consume_stock(memcg, nr_pages))
1956 if (!do_swap_account ||
1957 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1958 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
1960 if (do_swap_account)
1961 page_counter_uncharge(&memcg->memsw, batch);
1962 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1964 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1968 if (batch > nr_pages) {
1974 * Unlike in global OOM situations, memcg is not in a physical
1975 * memory shortage. Allow dying and OOM-killed tasks to
1976 * bypass the last charges so that they can exit quickly and
1977 * free their memory.
1979 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
1980 fatal_signal_pending(current) ||
1981 current->flags & PF_EXITING))
1984 if (unlikely(task_in_memcg_oom(current)))
1987 if (!(gfp_mask & __GFP_WAIT))
1990 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
1992 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
1993 gfp_mask, may_swap);
1995 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
1999 drain_all_stock(mem_over_limit);
2004 if (gfp_mask & __GFP_NORETRY)
2007 * Even though the limit is exceeded at this point, reclaim
2008 * may have been able to free some pages. Retry the charge
2009 * before killing the task.
2011 * Only for regular pages, though: huge pages are rather
2012 * unlikely to succeed so close to the limit, and we fall back
2013 * to regular pages anyway in case of failure.
2015 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2018 * At task move, charge accounts can be doubly counted. So, it's
2019 * better to wait until the end of task_move if something is going on.
2021 if (mem_cgroup_wait_acct_move(mem_over_limit))
2027 if (gfp_mask & __GFP_NOFAIL)
2030 if (fatal_signal_pending(current))
2033 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2035 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2037 if (!(gfp_mask & __GFP_NOFAIL))
2043 css_get_many(&memcg->css, batch);
2044 if (batch > nr_pages)
2045 refill_stock(memcg, batch - nr_pages);
2046 if (!(gfp_mask & __GFP_WAIT))
2049 * If the hierarchy is above the normal consumption range,
2050 * make the charging task trim their excess contribution.
2053 if (page_counter_read(&memcg->memory) <= memcg->high)
2055 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
2056 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2057 } while ((memcg = parent_mem_cgroup(memcg)));
2062 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2064 if (mem_cgroup_is_root(memcg))
2067 page_counter_uncharge(&memcg->memory, nr_pages);
2068 if (do_swap_account)
2069 page_counter_uncharge(&memcg->memsw, nr_pages);
2071 css_put_many(&memcg->css, nr_pages);
2075 * try_get_mem_cgroup_from_page - look up page's memcg association
2078 * Look up, get a css reference, and return the memcg that owns @page.
2080 * The page must be locked to prevent racing with swap-in and page
2081 * cache charges. If coming from an unlocked page table, the caller
2082 * must ensure the page is on the LRU or this can race with charging.
2084 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2086 struct mem_cgroup *memcg;
2090 VM_BUG_ON_PAGE(!PageLocked(page), page);
2092 memcg = page->mem_cgroup;
2094 if (!css_tryget_online(&memcg->css))
2096 } else if (PageSwapCache(page)) {
2097 ent.val = page_private(page);
2098 id = lookup_swap_cgroup_id(ent);
2100 memcg = mem_cgroup_from_id(id);
2101 if (memcg && !css_tryget_online(&memcg->css))
2108 static void lock_page_lru(struct page *page, int *isolated)
2110 struct zone *zone = page_zone(page);
2112 spin_lock_irq(&zone->lru_lock);
2113 if (PageLRU(page)) {
2114 struct lruvec *lruvec;
2116 lruvec = mem_cgroup_page_lruvec(page, zone);
2118 del_page_from_lru_list(page, lruvec, page_lru(page));
2124 static void unlock_page_lru(struct page *page, int isolated)
2126 struct zone *zone = page_zone(page);
2129 struct lruvec *lruvec;
2131 lruvec = mem_cgroup_page_lruvec(page, zone);
2132 VM_BUG_ON_PAGE(PageLRU(page), page);
2134 add_page_to_lru_list(page, lruvec, page_lru(page));
2136 spin_unlock_irq(&zone->lru_lock);
2139 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2144 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2147 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2148 * may already be on some other mem_cgroup's LRU. Take care of it.
2151 lock_page_lru(page, &isolated);
2154 * Nobody should be changing or seriously looking at
2155 * page->mem_cgroup at this point:
2157 * - the page is uncharged
2159 * - the page is off-LRU
2161 * - an anonymous fault has exclusive page access, except for
2162 * a locked page table
2164 * - a page cache insertion, a swapin fault, or a migration
2165 * have the page locked
2167 page->mem_cgroup = memcg;
2170 unlock_page_lru(page, isolated);
2173 #ifdef CONFIG_MEMCG_KMEM
2174 int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2175 unsigned long nr_pages)
2177 struct page_counter *counter;
2180 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2184 ret = try_charge(memcg, gfp, nr_pages);
2185 if (ret == -EINTR) {
2187 * try_charge() chose to bypass to root due to OOM kill or
2188 * fatal signal. Since our only options are to either fail
2189 * the allocation or charge it to this cgroup, do it as a
2190 * temporary condition. But we can't fail. From a kmem/slab
2191 * perspective, the cache has already been selected, by
2192 * mem_cgroup_kmem_get_cache(), so it is too late to change
2195 * This condition will only trigger if the task entered
2196 * memcg_charge_kmem in a sane state, but was OOM-killed
2197 * during try_charge() above. Tasks that were already dying
2198 * when the allocation triggers should have been already
2199 * directed to the root cgroup in memcontrol.h
2201 page_counter_charge(&memcg->memory, nr_pages);
2202 if (do_swap_account)
2203 page_counter_charge(&memcg->memsw, nr_pages);
2204 css_get_many(&memcg->css, nr_pages);
2207 page_counter_uncharge(&memcg->kmem, nr_pages);
2212 void memcg_uncharge_kmem(struct mem_cgroup *memcg, unsigned long nr_pages)
2214 page_counter_uncharge(&memcg->memory, nr_pages);
2215 if (do_swap_account)
2216 page_counter_uncharge(&memcg->memsw, nr_pages);
2218 page_counter_uncharge(&memcg->kmem, nr_pages);
2220 css_put_many(&memcg->css, nr_pages);
2223 static int memcg_alloc_cache_id(void)
2228 id = ida_simple_get(&memcg_cache_ida,
2229 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2233 if (id < memcg_nr_cache_ids)
2237 * There's no space for the new id in memcg_caches arrays,
2238 * so we have to grow them.
2240 down_write(&memcg_cache_ids_sem);
2242 size = 2 * (id + 1);
2243 if (size < MEMCG_CACHES_MIN_SIZE)
2244 size = MEMCG_CACHES_MIN_SIZE;
2245 else if (size > MEMCG_CACHES_MAX_SIZE)
2246 size = MEMCG_CACHES_MAX_SIZE;
2248 err = memcg_update_all_caches(size);
2250 err = memcg_update_all_list_lrus(size);
2252 memcg_nr_cache_ids = size;
2254 up_write(&memcg_cache_ids_sem);
2257 ida_simple_remove(&memcg_cache_ida, id);
2263 static void memcg_free_cache_id(int id)
2265 ida_simple_remove(&memcg_cache_ida, id);
2268 struct memcg_kmem_cache_create_work {
2269 struct mem_cgroup *memcg;
2270 struct kmem_cache *cachep;
2271 struct work_struct work;
2274 static void memcg_kmem_cache_create_func(struct work_struct *w)
2276 struct memcg_kmem_cache_create_work *cw =
2277 container_of(w, struct memcg_kmem_cache_create_work, work);
2278 struct mem_cgroup *memcg = cw->memcg;
2279 struct kmem_cache *cachep = cw->cachep;
2281 memcg_create_kmem_cache(memcg, cachep);
2283 css_put(&memcg->css);
2288 * Enqueue the creation of a per-memcg kmem_cache.
2290 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2291 struct kmem_cache *cachep)
2293 struct memcg_kmem_cache_create_work *cw;
2295 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2299 css_get(&memcg->css);
2302 cw->cachep = cachep;
2303 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2305 schedule_work(&cw->work);
2308 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2309 struct kmem_cache *cachep)
2312 * We need to stop accounting when we kmalloc, because if the
2313 * corresponding kmalloc cache is not yet created, the first allocation
2314 * in __memcg_schedule_kmem_cache_create will recurse.
2316 * However, it is better to enclose the whole function. Depending on
2317 * the debugging options enabled, INIT_WORK(), for instance, can
2318 * trigger an allocation. This too, will make us recurse. Because at
2319 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2320 * the safest choice is to do it like this, wrapping the whole function.
2322 current->memcg_kmem_skip_account = 1;
2323 __memcg_schedule_kmem_cache_create(memcg, cachep);
2324 current->memcg_kmem_skip_account = 0;
2328 * Return the kmem_cache we're supposed to use for a slab allocation.
2329 * We try to use the current memcg's version of the cache.
2331 * If the cache does not exist yet, if we are the first user of it,
2332 * we either create it immediately, if possible, or create it asynchronously
2334 * In the latter case, we will let the current allocation go through with
2335 * the original cache.
2337 * Can't be called in interrupt context or from kernel threads.
2338 * This function needs to be called with rcu_read_lock() held.
2340 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2342 struct mem_cgroup *memcg;
2343 struct kmem_cache *memcg_cachep;
2346 VM_BUG_ON(!is_root_cache(cachep));
2348 if (current->memcg_kmem_skip_account)
2351 memcg = get_mem_cgroup_from_mm(current->mm);
2352 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2356 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2357 if (likely(memcg_cachep))
2358 return memcg_cachep;
2361 * If we are in a safe context (can wait, and not in interrupt
2362 * context), we could be be predictable and return right away.
2363 * This would guarantee that the allocation being performed
2364 * already belongs in the new cache.
2366 * However, there are some clashes that can arrive from locking.
2367 * For instance, because we acquire the slab_mutex while doing
2368 * memcg_create_kmem_cache, this means no further allocation
2369 * could happen with the slab_mutex held. So it's better to
2372 memcg_schedule_kmem_cache_create(memcg, cachep);
2374 css_put(&memcg->css);
2378 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2380 if (!is_root_cache(cachep))
2381 css_put(&cachep->memcg_params.memcg->css);
2385 * We need to verify if the allocation against current->mm->owner's memcg is
2386 * possible for the given order. But the page is not allocated yet, so we'll
2387 * need a further commit step to do the final arrangements.
2389 * It is possible for the task to switch cgroups in this mean time, so at
2390 * commit time, we can't rely on task conversion any longer. We'll then use
2391 * the handle argument to return to the caller which cgroup we should commit
2392 * against. We could also return the memcg directly and avoid the pointer
2393 * passing, but a boolean return value gives better semantics considering
2394 * the compiled-out case as well.
2396 * Returning true means the allocation is possible.
2399 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2401 struct mem_cgroup *memcg;
2406 memcg = get_mem_cgroup_from_mm(current->mm);
2408 if (!memcg_kmem_is_active(memcg)) {
2409 css_put(&memcg->css);
2413 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2417 css_put(&memcg->css);
2421 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2424 VM_BUG_ON(mem_cgroup_is_root(memcg));
2426 /* The page allocation failed. Revert */
2428 memcg_uncharge_kmem(memcg, 1 << order);
2431 page->mem_cgroup = memcg;
2434 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2436 struct mem_cgroup *memcg = page->mem_cgroup;
2441 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2443 memcg_uncharge_kmem(memcg, 1 << order);
2444 page->mem_cgroup = NULL;
2447 struct mem_cgroup *__mem_cgroup_from_kmem(void *ptr)
2449 struct mem_cgroup *memcg = NULL;
2450 struct kmem_cache *cachep;
2453 page = virt_to_head_page(ptr);
2454 if (PageSlab(page)) {
2455 cachep = page->slab_cache;
2456 if (!is_root_cache(cachep))
2457 memcg = cachep->memcg_params.memcg;
2459 /* page allocated by alloc_kmem_pages */
2460 memcg = page->mem_cgroup;
2464 #endif /* CONFIG_MEMCG_KMEM */
2466 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2469 * Because tail pages are not marked as "used", set it. We're under
2470 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2471 * charge/uncharge will be never happen and move_account() is done under
2472 * compound_lock(), so we don't have to take care of races.
2474 void mem_cgroup_split_huge_fixup(struct page *head)
2478 if (mem_cgroup_disabled())
2481 for (i = 1; i < HPAGE_PMD_NR; i++)
2482 head[i].mem_cgroup = head->mem_cgroup;
2484 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2487 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2489 #ifdef CONFIG_MEMCG_SWAP
2490 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2493 int val = (charge) ? 1 : -1;
2494 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2498 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2499 * @entry: swap entry to be moved
2500 * @from: mem_cgroup which the entry is moved from
2501 * @to: mem_cgroup which the entry is moved to
2503 * It succeeds only when the swap_cgroup's record for this entry is the same
2504 * as the mem_cgroup's id of @from.
2506 * Returns 0 on success, -EINVAL on failure.
2508 * The caller must have charged to @to, IOW, called page_counter_charge() about
2509 * both res and memsw, and called css_get().
2511 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2512 struct mem_cgroup *from, struct mem_cgroup *to)
2514 unsigned short old_id, new_id;
2516 old_id = mem_cgroup_id(from);
2517 new_id = mem_cgroup_id(to);
2519 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2520 mem_cgroup_swap_statistics(from, false);
2521 mem_cgroup_swap_statistics(to, true);
2527 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2528 struct mem_cgroup *from, struct mem_cgroup *to)
2534 static DEFINE_MUTEX(memcg_limit_mutex);
2536 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2537 unsigned long limit)
2539 unsigned long curusage;
2540 unsigned long oldusage;
2541 bool enlarge = false;
2546 * For keeping hierarchical_reclaim simple, how long we should retry
2547 * is depends on callers. We set our retry-count to be function
2548 * of # of children which we should visit in this loop.
2550 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2551 mem_cgroup_count_children(memcg);
2553 oldusage = page_counter_read(&memcg->memory);
2556 if (signal_pending(current)) {
2561 mutex_lock(&memcg_limit_mutex);
2562 if (limit > memcg->memsw.limit) {
2563 mutex_unlock(&memcg_limit_mutex);
2567 if (limit > memcg->memory.limit)
2569 ret = page_counter_limit(&memcg->memory, limit);
2570 mutex_unlock(&memcg_limit_mutex);
2575 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2577 curusage = page_counter_read(&memcg->memory);
2578 /* Usage is reduced ? */
2579 if (curusage >= oldusage)
2582 oldusage = curusage;
2583 } while (retry_count);
2585 if (!ret && enlarge)
2586 memcg_oom_recover(memcg);
2591 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2592 unsigned long limit)
2594 unsigned long curusage;
2595 unsigned long oldusage;
2596 bool enlarge = false;
2600 /* see mem_cgroup_resize_res_limit */
2601 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2602 mem_cgroup_count_children(memcg);
2604 oldusage = page_counter_read(&memcg->memsw);
2607 if (signal_pending(current)) {
2612 mutex_lock(&memcg_limit_mutex);
2613 if (limit < memcg->memory.limit) {
2614 mutex_unlock(&memcg_limit_mutex);
2618 if (limit > memcg->memsw.limit)
2620 ret = page_counter_limit(&memcg->memsw, limit);
2621 mutex_unlock(&memcg_limit_mutex);
2626 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2628 curusage = page_counter_read(&memcg->memsw);
2629 /* Usage is reduced ? */
2630 if (curusage >= oldusage)
2633 oldusage = curusage;
2634 } while (retry_count);
2636 if (!ret && enlarge)
2637 memcg_oom_recover(memcg);
2642 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2644 unsigned long *total_scanned)
2646 unsigned long nr_reclaimed = 0;
2647 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2648 unsigned long reclaimed;
2650 struct mem_cgroup_tree_per_zone *mctz;
2651 unsigned long excess;
2652 unsigned long nr_scanned;
2657 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2659 * This loop can run a while, specially if mem_cgroup's continuously
2660 * keep exceeding their soft limit and putting the system under
2667 mz = mem_cgroup_largest_soft_limit_node(mctz);
2672 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2673 gfp_mask, &nr_scanned);
2674 nr_reclaimed += reclaimed;
2675 *total_scanned += nr_scanned;
2676 spin_lock_irq(&mctz->lock);
2677 __mem_cgroup_remove_exceeded(mz, mctz);
2680 * If we failed to reclaim anything from this memory cgroup
2681 * it is time to move on to the next cgroup
2685 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2687 excess = soft_limit_excess(mz->memcg);
2689 * One school of thought says that we should not add
2690 * back the node to the tree if reclaim returns 0.
2691 * But our reclaim could return 0, simply because due
2692 * to priority we are exposing a smaller subset of
2693 * memory to reclaim from. Consider this as a longer
2696 /* If excess == 0, no tree ops */
2697 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2698 spin_unlock_irq(&mctz->lock);
2699 css_put(&mz->memcg->css);
2702 * Could not reclaim anything and there are no more
2703 * mem cgroups to try or we seem to be looping without
2704 * reclaiming anything.
2706 if (!nr_reclaimed &&
2708 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2710 } while (!nr_reclaimed);
2712 css_put(&next_mz->memcg->css);
2713 return nr_reclaimed;
2717 * Test whether @memcg has children, dead or alive. Note that this
2718 * function doesn't care whether @memcg has use_hierarchy enabled and
2719 * returns %true if there are child csses according to the cgroup
2720 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2722 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2727 * The lock does not prevent addition or deletion of children, but
2728 * it prevents a new child from being initialized based on this
2729 * parent in css_online(), so it's enough to decide whether
2730 * hierarchically inherited attributes can still be changed or not.
2732 lockdep_assert_held(&memcg_create_mutex);
2735 ret = css_next_child(NULL, &memcg->css);
2741 * Reclaims as many pages from the given memcg as possible and moves
2742 * the rest to the parent.
2744 * Caller is responsible for holding css reference for memcg.
2746 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2748 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2750 /* we call try-to-free pages for make this cgroup empty */
2751 lru_add_drain_all();
2752 /* try to free all pages in this cgroup */
2753 while (nr_retries && page_counter_read(&memcg->memory)) {
2756 if (signal_pending(current))
2759 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2763 /* maybe some writeback is necessary */
2764 congestion_wait(BLK_RW_ASYNC, HZ/10);
2772 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2773 char *buf, size_t nbytes,
2776 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2778 if (mem_cgroup_is_root(memcg))
2780 return mem_cgroup_force_empty(memcg) ?: nbytes;
2783 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2786 return mem_cgroup_from_css(css)->use_hierarchy;
2789 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2790 struct cftype *cft, u64 val)
2793 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2794 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2796 mutex_lock(&memcg_create_mutex);
2798 if (memcg->use_hierarchy == val)
2802 * If parent's use_hierarchy is set, we can't make any modifications
2803 * in the child subtrees. If it is unset, then the change can
2804 * occur, provided the current cgroup has no children.
2806 * For the root cgroup, parent_mem is NULL, we allow value to be
2807 * set if there are no children.
2809 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2810 (val == 1 || val == 0)) {
2811 if (!memcg_has_children(memcg))
2812 memcg->use_hierarchy = val;
2819 mutex_unlock(&memcg_create_mutex);
2824 static unsigned long tree_stat(struct mem_cgroup *memcg,
2825 enum mem_cgroup_stat_index idx)
2827 struct mem_cgroup *iter;
2830 /* Per-cpu values can be negative, use a signed accumulator */
2831 for_each_mem_cgroup_tree(iter, memcg)
2832 val += mem_cgroup_read_stat(iter, idx);
2834 if (val < 0) /* race ? */
2839 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2843 if (mem_cgroup_is_root(memcg)) {
2844 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
2845 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
2847 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
2850 val = page_counter_read(&memcg->memory);
2852 val = page_counter_read(&memcg->memsw);
2854 return val << PAGE_SHIFT;
2865 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2868 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2869 struct page_counter *counter;
2871 switch (MEMFILE_TYPE(cft->private)) {
2873 counter = &memcg->memory;
2876 counter = &memcg->memsw;
2879 counter = &memcg->kmem;
2885 switch (MEMFILE_ATTR(cft->private)) {
2887 if (counter == &memcg->memory)
2888 return mem_cgroup_usage(memcg, false);
2889 if (counter == &memcg->memsw)
2890 return mem_cgroup_usage(memcg, true);
2891 return (u64)page_counter_read(counter) * PAGE_SIZE;
2893 return (u64)counter->limit * PAGE_SIZE;
2895 return (u64)counter->watermark * PAGE_SIZE;
2897 return counter->failcnt;
2898 case RES_SOFT_LIMIT:
2899 return (u64)memcg->soft_limit * PAGE_SIZE;
2905 #ifdef CONFIG_MEMCG_KMEM
2906 static int memcg_activate_kmem(struct mem_cgroup *memcg,
2907 unsigned long nr_pages)
2912 BUG_ON(memcg->kmemcg_id >= 0);
2913 BUG_ON(memcg->kmem_acct_activated);
2914 BUG_ON(memcg->kmem_acct_active);
2917 * For simplicity, we won't allow this to be disabled. It also can't
2918 * be changed if the cgroup has children already, or if tasks had
2921 * If tasks join before we set the limit, a person looking at
2922 * kmem.usage_in_bytes will have no way to determine when it took
2923 * place, which makes the value quite meaningless.
2925 * After it first became limited, changes in the value of the limit are
2926 * of course permitted.
2928 mutex_lock(&memcg_create_mutex);
2929 if (cgroup_has_tasks(memcg->css.cgroup) ||
2930 (memcg->use_hierarchy && memcg_has_children(memcg)))
2932 mutex_unlock(&memcg_create_mutex);
2936 memcg_id = memcg_alloc_cache_id();
2943 * We couldn't have accounted to this cgroup, because it hasn't got
2944 * activated yet, so this should succeed.
2946 err = page_counter_limit(&memcg->kmem, nr_pages);
2949 static_key_slow_inc(&memcg_kmem_enabled_key);
2951 * A memory cgroup is considered kmem-active as soon as it gets
2952 * kmemcg_id. Setting the id after enabling static branching will
2953 * guarantee no one starts accounting before all call sites are
2956 memcg->kmemcg_id = memcg_id;
2957 memcg->kmem_acct_activated = true;
2958 memcg->kmem_acct_active = true;
2963 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2964 unsigned long limit)
2968 mutex_lock(&memcg_limit_mutex);
2969 if (!memcg_kmem_is_active(memcg))
2970 ret = memcg_activate_kmem(memcg, limit);
2972 ret = page_counter_limit(&memcg->kmem, limit);
2973 mutex_unlock(&memcg_limit_mutex);
2977 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
2980 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
2985 mutex_lock(&memcg_limit_mutex);
2987 * If the parent cgroup is not kmem-active now, it cannot be activated
2988 * after this point, because it has at least one child already.
2990 if (memcg_kmem_is_active(parent))
2991 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
2992 mutex_unlock(&memcg_limit_mutex);
2996 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2997 unsigned long limit)
3001 #endif /* CONFIG_MEMCG_KMEM */
3004 * The user of this function is...
3007 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3008 char *buf, size_t nbytes, loff_t off)
3010 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3011 unsigned long nr_pages;
3014 buf = strstrip(buf);
3015 ret = page_counter_memparse(buf, "-1", &nr_pages);
3019 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3021 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3025 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3027 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3030 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3033 ret = memcg_update_kmem_limit(memcg, nr_pages);
3037 case RES_SOFT_LIMIT:
3038 memcg->soft_limit = nr_pages;
3042 return ret ?: nbytes;
3045 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3046 size_t nbytes, loff_t off)
3048 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3049 struct page_counter *counter;
3051 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3053 counter = &memcg->memory;
3056 counter = &memcg->memsw;
3059 counter = &memcg->kmem;
3065 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3067 page_counter_reset_watermark(counter);
3070 counter->failcnt = 0;
3079 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3082 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3086 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3087 struct cftype *cft, u64 val)
3089 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3091 if (val & ~MOVE_MASK)
3095 * No kind of locking is needed in here, because ->can_attach() will
3096 * check this value once in the beginning of the process, and then carry
3097 * on with stale data. This means that changes to this value will only
3098 * affect task migrations starting after the change.
3100 memcg->move_charge_at_immigrate = val;
3104 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3105 struct cftype *cft, u64 val)
3112 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3116 unsigned int lru_mask;
3119 static const struct numa_stat stats[] = {
3120 { "total", LRU_ALL },
3121 { "file", LRU_ALL_FILE },
3122 { "anon", LRU_ALL_ANON },
3123 { "unevictable", BIT(LRU_UNEVICTABLE) },
3125 const struct numa_stat *stat;
3128 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3130 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3131 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3132 seq_printf(m, "%s=%lu", stat->name, nr);
3133 for_each_node_state(nid, N_MEMORY) {
3134 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3136 seq_printf(m, " N%d=%lu", nid, nr);
3141 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3142 struct mem_cgroup *iter;
3145 for_each_mem_cgroup_tree(iter, memcg)
3146 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3147 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3148 for_each_node_state(nid, N_MEMORY) {
3150 for_each_mem_cgroup_tree(iter, memcg)
3151 nr += mem_cgroup_node_nr_lru_pages(
3152 iter, nid, stat->lru_mask);
3153 seq_printf(m, " N%d=%lu", nid, nr);
3160 #endif /* CONFIG_NUMA */
3162 static int memcg_stat_show(struct seq_file *m, void *v)
3164 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3165 unsigned long memory, memsw;
3166 struct mem_cgroup *mi;
3169 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3170 MEM_CGROUP_STAT_NSTATS);
3171 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3172 MEM_CGROUP_EVENTS_NSTATS);
3173 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3175 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3176 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3178 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3179 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3182 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3183 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3184 mem_cgroup_read_events(memcg, i));
3186 for (i = 0; i < NR_LRU_LISTS; i++)
3187 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3188 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3190 /* Hierarchical information */
3191 memory = memsw = PAGE_COUNTER_MAX;
3192 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3193 memory = min(memory, mi->memory.limit);
3194 memsw = min(memsw, mi->memsw.limit);
3196 seq_printf(m, "hierarchical_memory_limit %llu\n",
3197 (u64)memory * PAGE_SIZE);
3198 if (do_swap_account)
3199 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3200 (u64)memsw * PAGE_SIZE);
3202 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3205 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3207 for_each_mem_cgroup_tree(mi, memcg)
3208 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3209 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3212 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3213 unsigned long long val = 0;
3215 for_each_mem_cgroup_tree(mi, memcg)
3216 val += mem_cgroup_read_events(mi, i);
3217 seq_printf(m, "total_%s %llu\n",
3218 mem_cgroup_events_names[i], val);
3221 for (i = 0; i < NR_LRU_LISTS; i++) {
3222 unsigned long long val = 0;
3224 for_each_mem_cgroup_tree(mi, memcg)
3225 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3226 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3229 #ifdef CONFIG_DEBUG_VM
3232 struct mem_cgroup_per_zone *mz;
3233 struct zone_reclaim_stat *rstat;
3234 unsigned long recent_rotated[2] = {0, 0};
3235 unsigned long recent_scanned[2] = {0, 0};
3237 for_each_online_node(nid)
3238 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3239 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3240 rstat = &mz->lruvec.reclaim_stat;
3242 recent_rotated[0] += rstat->recent_rotated[0];
3243 recent_rotated[1] += rstat->recent_rotated[1];
3244 recent_scanned[0] += rstat->recent_scanned[0];
3245 recent_scanned[1] += rstat->recent_scanned[1];
3247 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3248 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3249 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3250 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3257 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3260 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3262 return mem_cgroup_swappiness(memcg);
3265 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3266 struct cftype *cft, u64 val)
3268 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3274 memcg->swappiness = val;
3276 vm_swappiness = val;
3281 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3283 struct mem_cgroup_threshold_ary *t;
3284 unsigned long usage;
3289 t = rcu_dereference(memcg->thresholds.primary);
3291 t = rcu_dereference(memcg->memsw_thresholds.primary);
3296 usage = mem_cgroup_usage(memcg, swap);
3299 * current_threshold points to threshold just below or equal to usage.
3300 * If it's not true, a threshold was crossed after last
3301 * call of __mem_cgroup_threshold().
3303 i = t->current_threshold;
3306 * Iterate backward over array of thresholds starting from
3307 * current_threshold and check if a threshold is crossed.
3308 * If none of thresholds below usage is crossed, we read
3309 * only one element of the array here.
3311 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3312 eventfd_signal(t->entries[i].eventfd, 1);
3314 /* i = current_threshold + 1 */
3318 * Iterate forward over array of thresholds starting from
3319 * current_threshold+1 and check if a threshold is crossed.
3320 * If none of thresholds above usage is crossed, we read
3321 * only one element of the array here.
3323 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3324 eventfd_signal(t->entries[i].eventfd, 1);
3326 /* Update current_threshold */
3327 t->current_threshold = i - 1;
3332 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3335 __mem_cgroup_threshold(memcg, false);
3336 if (do_swap_account)
3337 __mem_cgroup_threshold(memcg, true);
3339 memcg = parent_mem_cgroup(memcg);
3343 static int compare_thresholds(const void *a, const void *b)
3345 const struct mem_cgroup_threshold *_a = a;
3346 const struct mem_cgroup_threshold *_b = b;
3348 if (_a->threshold > _b->threshold)
3351 if (_a->threshold < _b->threshold)
3357 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3359 struct mem_cgroup_eventfd_list *ev;
3361 spin_lock(&memcg_oom_lock);
3363 list_for_each_entry(ev, &memcg->oom_notify, list)
3364 eventfd_signal(ev->eventfd, 1);
3366 spin_unlock(&memcg_oom_lock);
3370 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3372 struct mem_cgroup *iter;
3374 for_each_mem_cgroup_tree(iter, memcg)
3375 mem_cgroup_oom_notify_cb(iter);
3378 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3379 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3381 struct mem_cgroup_thresholds *thresholds;
3382 struct mem_cgroup_threshold_ary *new;
3383 unsigned long threshold;
3384 unsigned long usage;
3387 ret = page_counter_memparse(args, "-1", &threshold);
3391 mutex_lock(&memcg->thresholds_lock);
3394 thresholds = &memcg->thresholds;
3395 usage = mem_cgroup_usage(memcg, false);
3396 } else if (type == _MEMSWAP) {
3397 thresholds = &memcg->memsw_thresholds;
3398 usage = mem_cgroup_usage(memcg, true);
3402 /* Check if a threshold crossed before adding a new one */
3403 if (thresholds->primary)
3404 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3406 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3408 /* Allocate memory for new array of thresholds */
3409 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3417 /* Copy thresholds (if any) to new array */
3418 if (thresholds->primary) {
3419 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3420 sizeof(struct mem_cgroup_threshold));
3423 /* Add new threshold */
3424 new->entries[size - 1].eventfd = eventfd;
3425 new->entries[size - 1].threshold = threshold;
3427 /* Sort thresholds. Registering of new threshold isn't time-critical */
3428 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3429 compare_thresholds, NULL);
3431 /* Find current threshold */
3432 new->current_threshold = -1;
3433 for (i = 0; i < size; i++) {
3434 if (new->entries[i].threshold <= usage) {
3436 * new->current_threshold will not be used until
3437 * rcu_assign_pointer(), so it's safe to increment
3440 ++new->current_threshold;
3445 /* Free old spare buffer and save old primary buffer as spare */
3446 kfree(thresholds->spare);
3447 thresholds->spare = thresholds->primary;
3449 rcu_assign_pointer(thresholds->primary, new);
3451 /* To be sure that nobody uses thresholds */
3455 mutex_unlock(&memcg->thresholds_lock);
3460 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3461 struct eventfd_ctx *eventfd, const char *args)
3463 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3466 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3467 struct eventfd_ctx *eventfd, const char *args)
3469 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3472 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3473 struct eventfd_ctx *eventfd, enum res_type type)
3475 struct mem_cgroup_thresholds *thresholds;
3476 struct mem_cgroup_threshold_ary *new;
3477 unsigned long usage;
3480 mutex_lock(&memcg->thresholds_lock);
3483 thresholds = &memcg->thresholds;
3484 usage = mem_cgroup_usage(memcg, false);
3485 } else if (type == _MEMSWAP) {
3486 thresholds = &memcg->memsw_thresholds;
3487 usage = mem_cgroup_usage(memcg, true);
3491 if (!thresholds->primary)
3494 /* Check if a threshold crossed before removing */
3495 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3497 /* Calculate new number of threshold */
3499 for (i = 0; i < thresholds->primary->size; i++) {
3500 if (thresholds->primary->entries[i].eventfd != eventfd)
3504 new = thresholds->spare;
3506 /* Set thresholds array to NULL if we don't have thresholds */
3515 /* Copy thresholds and find current threshold */
3516 new->current_threshold = -1;
3517 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3518 if (thresholds->primary->entries[i].eventfd == eventfd)
3521 new->entries[j] = thresholds->primary->entries[i];
3522 if (new->entries[j].threshold <= usage) {
3524 * new->current_threshold will not be used
3525 * until rcu_assign_pointer(), so it's safe to increment
3528 ++new->current_threshold;
3534 /* Swap primary and spare array */
3535 thresholds->spare = thresholds->primary;
3536 /* If all events are unregistered, free the spare array */
3538 kfree(thresholds->spare);
3539 thresholds->spare = NULL;
3542 rcu_assign_pointer(thresholds->primary, new);
3544 /* To be sure that nobody uses thresholds */
3547 mutex_unlock(&memcg->thresholds_lock);
3550 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3551 struct eventfd_ctx *eventfd)
3553 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3556 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3557 struct eventfd_ctx *eventfd)
3559 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3562 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3563 struct eventfd_ctx *eventfd, const char *args)
3565 struct mem_cgroup_eventfd_list *event;
3567 event = kmalloc(sizeof(*event), GFP_KERNEL);
3571 spin_lock(&memcg_oom_lock);
3573 event->eventfd = eventfd;
3574 list_add(&event->list, &memcg->oom_notify);
3576 /* already in OOM ? */
3577 if (memcg->under_oom)
3578 eventfd_signal(eventfd, 1);
3579 spin_unlock(&memcg_oom_lock);
3584 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3585 struct eventfd_ctx *eventfd)
3587 struct mem_cgroup_eventfd_list *ev, *tmp;
3589 spin_lock(&memcg_oom_lock);
3591 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3592 if (ev->eventfd == eventfd) {
3593 list_del(&ev->list);
3598 spin_unlock(&memcg_oom_lock);
3601 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3603 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3605 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3606 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3610 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3611 struct cftype *cft, u64 val)
3613 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3615 /* cannot set to root cgroup and only 0 and 1 are allowed */
3616 if (!css->parent || !((val == 0) || (val == 1)))
3619 memcg->oom_kill_disable = val;
3621 memcg_oom_recover(memcg);
3626 #ifdef CONFIG_MEMCG_KMEM
3627 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3631 ret = memcg_propagate_kmem(memcg);
3635 return mem_cgroup_sockets_init(memcg, ss);
3638 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3640 struct cgroup_subsys_state *css;
3641 struct mem_cgroup *parent, *child;
3644 if (!memcg->kmem_acct_active)
3648 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3649 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3650 * guarantees no cache will be created for this cgroup after we are
3651 * done (see memcg_create_kmem_cache()).
3653 memcg->kmem_acct_active = false;
3655 memcg_deactivate_kmem_caches(memcg);
3657 kmemcg_id = memcg->kmemcg_id;
3658 BUG_ON(kmemcg_id < 0);
3660 parent = parent_mem_cgroup(memcg);
3662 parent = root_mem_cgroup;
3665 * Change kmemcg_id of this cgroup and all its descendants to the
3666 * parent's id, and then move all entries from this cgroup's list_lrus
3667 * to ones of the parent. After we have finished, all list_lrus
3668 * corresponding to this cgroup are guaranteed to remain empty. The
3669 * ordering is imposed by list_lru_node->lock taken by
3670 * memcg_drain_all_list_lrus().
3672 css_for_each_descendant_pre(css, &memcg->css) {
3673 child = mem_cgroup_from_css(css);
3674 BUG_ON(child->kmemcg_id != kmemcg_id);
3675 child->kmemcg_id = parent->kmemcg_id;
3676 if (!memcg->use_hierarchy)
3679 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3681 memcg_free_cache_id(kmemcg_id);
3684 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3686 if (memcg->kmem_acct_activated) {
3687 memcg_destroy_kmem_caches(memcg);
3688 static_key_slow_dec(&memcg_kmem_enabled_key);
3689 WARN_ON(page_counter_read(&memcg->kmem));
3691 mem_cgroup_sockets_destroy(memcg);
3694 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3699 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3703 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3708 #ifdef CONFIG_CGROUP_WRITEBACK
3710 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3712 return &memcg->cgwb_list;
3715 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3717 return wb_domain_init(&memcg->cgwb_domain, gfp);
3720 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3722 wb_domain_exit(&memcg->cgwb_domain);
3725 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3727 wb_domain_size_changed(&memcg->cgwb_domain);
3730 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3732 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3734 if (!memcg->css.parent)
3737 return &memcg->cgwb_domain;
3741 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3742 * @wb: bdi_writeback in question
3743 * @pavail: out parameter for number of available pages
3744 * @pdirty: out parameter for number of dirty pages
3745 * @pwriteback: out parameter for number of pages under writeback
3747 * Determine the numbers of available, dirty, and writeback pages in @wb's
3748 * memcg. Dirty and writeback are self-explanatory. Available is a bit
3751 * A memcg's headroom is "min(max, high) - used". The available memory is
3752 * calculated as the lowest headroom of itself and the ancestors plus the
3753 * number of pages already being used for file pages. Note that this
3754 * doesn't consider the actual amount of available memory in the system.
3755 * The caller should further cap *@pavail accordingly.
3757 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pavail,
3758 unsigned long *pdirty, unsigned long *pwriteback)
3760 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3761 struct mem_cgroup *parent;
3762 unsigned long head_room = PAGE_COUNTER_MAX;
3763 unsigned long file_pages;
3765 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3767 /* this should eventually include NR_UNSTABLE_NFS */
3768 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3770 file_pages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3771 (1 << LRU_ACTIVE_FILE));
3772 while ((parent = parent_mem_cgroup(memcg))) {
3773 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3774 unsigned long used = page_counter_read(&memcg->memory);
3776 head_room = min(head_room, ceiling - min(ceiling, used));
3780 *pavail = file_pages + head_room;
3783 #else /* CONFIG_CGROUP_WRITEBACK */
3785 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3790 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3794 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3798 #endif /* CONFIG_CGROUP_WRITEBACK */
3801 * DO NOT USE IN NEW FILES.
3803 * "cgroup.event_control" implementation.
3805 * This is way over-engineered. It tries to support fully configurable
3806 * events for each user. Such level of flexibility is completely
3807 * unnecessary especially in the light of the planned unified hierarchy.
3809 * Please deprecate this and replace with something simpler if at all
3814 * Unregister event and free resources.
3816 * Gets called from workqueue.
3818 static void memcg_event_remove(struct work_struct *work)
3820 struct mem_cgroup_event *event =
3821 container_of(work, struct mem_cgroup_event, remove);
3822 struct mem_cgroup *memcg = event->memcg;
3824 remove_wait_queue(event->wqh, &event->wait);
3826 event->unregister_event(memcg, event->eventfd);
3828 /* Notify userspace the event is going away. */
3829 eventfd_signal(event->eventfd, 1);
3831 eventfd_ctx_put(event->eventfd);
3833 css_put(&memcg->css);
3837 * Gets called on POLLHUP on eventfd when user closes it.
3839 * Called with wqh->lock held and interrupts disabled.
3841 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3842 int sync, void *key)
3844 struct mem_cgroup_event *event =
3845 container_of(wait, struct mem_cgroup_event, wait);
3846 struct mem_cgroup *memcg = event->memcg;
3847 unsigned long flags = (unsigned long)key;
3849 if (flags & POLLHUP) {
3851 * If the event has been detached at cgroup removal, we
3852 * can simply return knowing the other side will cleanup
3855 * We can't race against event freeing since the other
3856 * side will require wqh->lock via remove_wait_queue(),
3859 spin_lock(&memcg->event_list_lock);
3860 if (!list_empty(&event->list)) {
3861 list_del_init(&event->list);
3863 * We are in atomic context, but cgroup_event_remove()
3864 * may sleep, so we have to call it in workqueue.
3866 schedule_work(&event->remove);
3868 spin_unlock(&memcg->event_list_lock);
3874 static void memcg_event_ptable_queue_proc(struct file *file,
3875 wait_queue_head_t *wqh, poll_table *pt)
3877 struct mem_cgroup_event *event =
3878 container_of(pt, struct mem_cgroup_event, pt);
3881 add_wait_queue(wqh, &event->wait);
3885 * DO NOT USE IN NEW FILES.
3887 * Parse input and register new cgroup event handler.
3889 * Input must be in format '<event_fd> <control_fd> <args>'.
3890 * Interpretation of args is defined by control file implementation.
3892 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3893 char *buf, size_t nbytes, loff_t off)
3895 struct cgroup_subsys_state *css = of_css(of);
3896 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3897 struct mem_cgroup_event *event;
3898 struct cgroup_subsys_state *cfile_css;
3899 unsigned int efd, cfd;
3906 buf = strstrip(buf);
3908 efd = simple_strtoul(buf, &endp, 10);
3913 cfd = simple_strtoul(buf, &endp, 10);
3914 if ((*endp != ' ') && (*endp != '\0'))
3918 event = kzalloc(sizeof(*event), GFP_KERNEL);
3922 event->memcg = memcg;
3923 INIT_LIST_HEAD(&event->list);
3924 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3925 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3926 INIT_WORK(&event->remove, memcg_event_remove);
3934 event->eventfd = eventfd_ctx_fileget(efile.file);
3935 if (IS_ERR(event->eventfd)) {
3936 ret = PTR_ERR(event->eventfd);
3943 goto out_put_eventfd;
3946 /* the process need read permission on control file */
3947 /* AV: shouldn't we check that it's been opened for read instead? */
3948 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3953 * Determine the event callbacks and set them in @event. This used
3954 * to be done via struct cftype but cgroup core no longer knows
3955 * about these events. The following is crude but the whole thing
3956 * is for compatibility anyway.
3958 * DO NOT ADD NEW FILES.
3960 name = cfile.file->f_path.dentry->d_name.name;
3962 if (!strcmp(name, "memory.usage_in_bytes")) {
3963 event->register_event = mem_cgroup_usage_register_event;
3964 event->unregister_event = mem_cgroup_usage_unregister_event;
3965 } else if (!strcmp(name, "memory.oom_control")) {
3966 event->register_event = mem_cgroup_oom_register_event;
3967 event->unregister_event = mem_cgroup_oom_unregister_event;
3968 } else if (!strcmp(name, "memory.pressure_level")) {
3969 event->register_event = vmpressure_register_event;
3970 event->unregister_event = vmpressure_unregister_event;
3971 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3972 event->register_event = memsw_cgroup_usage_register_event;
3973 event->unregister_event = memsw_cgroup_usage_unregister_event;
3980 * Verify @cfile should belong to @css. Also, remaining events are
3981 * automatically removed on cgroup destruction but the removal is
3982 * asynchronous, so take an extra ref on @css.
3984 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3985 &memory_cgrp_subsys);
3987 if (IS_ERR(cfile_css))
3989 if (cfile_css != css) {
3994 ret = event->register_event(memcg, event->eventfd, buf);
3998 efile.file->f_op->poll(efile.file, &event->pt);
4000 spin_lock(&memcg->event_list_lock);
4001 list_add(&event->list, &memcg->event_list);
4002 spin_unlock(&memcg->event_list_lock);
4014 eventfd_ctx_put(event->eventfd);
4023 static struct cftype mem_cgroup_legacy_files[] = {
4025 .name = "usage_in_bytes",
4026 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4027 .read_u64 = mem_cgroup_read_u64,
4030 .name = "max_usage_in_bytes",
4031 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4032 .write = mem_cgroup_reset,
4033 .read_u64 = mem_cgroup_read_u64,
4036 .name = "limit_in_bytes",
4037 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4038 .write = mem_cgroup_write,
4039 .read_u64 = mem_cgroup_read_u64,
4042 .name = "soft_limit_in_bytes",
4043 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4044 .write = mem_cgroup_write,
4045 .read_u64 = mem_cgroup_read_u64,
4049 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4050 .write = mem_cgroup_reset,
4051 .read_u64 = mem_cgroup_read_u64,
4055 .seq_show = memcg_stat_show,
4058 .name = "force_empty",
4059 .write = mem_cgroup_force_empty_write,
4062 .name = "use_hierarchy",
4063 .write_u64 = mem_cgroup_hierarchy_write,
4064 .read_u64 = mem_cgroup_hierarchy_read,
4067 .name = "cgroup.event_control", /* XXX: for compat */
4068 .write = memcg_write_event_control,
4069 .flags = CFTYPE_NO_PREFIX,
4073 .name = "swappiness",
4074 .read_u64 = mem_cgroup_swappiness_read,
4075 .write_u64 = mem_cgroup_swappiness_write,
4078 .name = "move_charge_at_immigrate",
4079 .read_u64 = mem_cgroup_move_charge_read,
4080 .write_u64 = mem_cgroup_move_charge_write,
4083 .name = "oom_control",
4084 .seq_show = mem_cgroup_oom_control_read,
4085 .write_u64 = mem_cgroup_oom_control_write,
4086 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4089 .name = "pressure_level",
4093 .name = "numa_stat",
4094 .seq_show = memcg_numa_stat_show,
4097 #ifdef CONFIG_MEMCG_KMEM
4099 .name = "kmem.limit_in_bytes",
4100 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4101 .write = mem_cgroup_write,
4102 .read_u64 = mem_cgroup_read_u64,
4105 .name = "kmem.usage_in_bytes",
4106 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4107 .read_u64 = mem_cgroup_read_u64,
4110 .name = "kmem.failcnt",
4111 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4112 .write = mem_cgroup_reset,
4113 .read_u64 = mem_cgroup_read_u64,
4116 .name = "kmem.max_usage_in_bytes",
4117 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4118 .write = mem_cgroup_reset,
4119 .read_u64 = mem_cgroup_read_u64,
4121 #ifdef CONFIG_SLABINFO
4123 .name = "kmem.slabinfo",
4124 .seq_start = slab_start,
4125 .seq_next = slab_next,
4126 .seq_stop = slab_stop,
4127 .seq_show = memcg_slab_show,
4131 { }, /* terminate */
4134 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4136 struct mem_cgroup_per_node *pn;
4137 struct mem_cgroup_per_zone *mz;
4138 int zone, tmp = node;
4140 * This routine is called against possible nodes.
4141 * But it's BUG to call kmalloc() against offline node.
4143 * TODO: this routine can waste much memory for nodes which will
4144 * never be onlined. It's better to use memory hotplug callback
4147 if (!node_state(node, N_NORMAL_MEMORY))
4149 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4153 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4154 mz = &pn->zoneinfo[zone];
4155 lruvec_init(&mz->lruvec);
4156 mz->usage_in_excess = 0;
4157 mz->on_tree = false;
4160 memcg->nodeinfo[node] = pn;
4164 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4166 kfree(memcg->nodeinfo[node]);
4169 static struct mem_cgroup *mem_cgroup_alloc(void)
4171 struct mem_cgroup *memcg;
4174 size = sizeof(struct mem_cgroup);
4175 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4177 memcg = kzalloc(size, GFP_KERNEL);
4181 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4185 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4188 spin_lock_init(&memcg->pcp_counter_lock);
4192 free_percpu(memcg->stat);
4199 * At destroying mem_cgroup, references from swap_cgroup can remain.
4200 * (scanning all at force_empty is too costly...)
4202 * Instead of clearing all references at force_empty, we remember
4203 * the number of reference from swap_cgroup and free mem_cgroup when
4204 * it goes down to 0.
4206 * Removal of cgroup itself succeeds regardless of refs from swap.
4209 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4213 mem_cgroup_remove_from_trees(memcg);
4216 free_mem_cgroup_per_zone_info(memcg, node);
4218 free_percpu(memcg->stat);
4219 memcg_wb_domain_exit(memcg);
4224 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4226 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4228 if (!memcg->memory.parent)
4230 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4232 EXPORT_SYMBOL(parent_mem_cgroup);
4234 static struct cgroup_subsys_state * __ref
4235 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4237 struct mem_cgroup *memcg;
4238 long error = -ENOMEM;
4241 memcg = mem_cgroup_alloc();
4243 return ERR_PTR(error);
4246 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4250 if (parent_css == NULL) {
4251 root_mem_cgroup = memcg;
4252 mem_cgroup_root_css = &memcg->css;
4253 page_counter_init(&memcg->memory, NULL);
4254 memcg->high = PAGE_COUNTER_MAX;
4255 memcg->soft_limit = PAGE_COUNTER_MAX;
4256 page_counter_init(&memcg->memsw, NULL);
4257 page_counter_init(&memcg->kmem, NULL);
4260 memcg->last_scanned_node = MAX_NUMNODES;
4261 INIT_LIST_HEAD(&memcg->oom_notify);
4262 memcg->move_charge_at_immigrate = 0;
4263 mutex_init(&memcg->thresholds_lock);
4264 spin_lock_init(&memcg->move_lock);
4265 vmpressure_init(&memcg->vmpressure);
4266 INIT_LIST_HEAD(&memcg->event_list);
4267 spin_lock_init(&memcg->event_list_lock);
4268 #ifdef CONFIG_MEMCG_KMEM
4269 memcg->kmemcg_id = -1;
4271 #ifdef CONFIG_CGROUP_WRITEBACK
4272 INIT_LIST_HEAD(&memcg->cgwb_list);
4277 __mem_cgroup_free(memcg);
4278 return ERR_PTR(error);
4282 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4284 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4285 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4288 if (css->id > MEM_CGROUP_ID_MAX)
4294 mutex_lock(&memcg_create_mutex);
4296 memcg->use_hierarchy = parent->use_hierarchy;
4297 memcg->oom_kill_disable = parent->oom_kill_disable;
4298 memcg->swappiness = mem_cgroup_swappiness(parent);
4300 if (parent->use_hierarchy) {
4301 page_counter_init(&memcg->memory, &parent->memory);
4302 memcg->high = PAGE_COUNTER_MAX;
4303 memcg->soft_limit = PAGE_COUNTER_MAX;
4304 page_counter_init(&memcg->memsw, &parent->memsw);
4305 page_counter_init(&memcg->kmem, &parent->kmem);
4308 * No need to take a reference to the parent because cgroup
4309 * core guarantees its existence.
4312 page_counter_init(&memcg->memory, NULL);
4313 memcg->high = PAGE_COUNTER_MAX;
4314 memcg->soft_limit = PAGE_COUNTER_MAX;
4315 page_counter_init(&memcg->memsw, NULL);
4316 page_counter_init(&memcg->kmem, NULL);
4318 * Deeper hierachy with use_hierarchy == false doesn't make
4319 * much sense so let cgroup subsystem know about this
4320 * unfortunate state in our controller.
4322 if (parent != root_mem_cgroup)
4323 memory_cgrp_subsys.broken_hierarchy = true;
4325 mutex_unlock(&memcg_create_mutex);
4327 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4332 * Make sure the memcg is initialized: mem_cgroup_iter()
4333 * orders reading memcg->initialized against its callers
4334 * reading the memcg members.
4336 smp_store_release(&memcg->initialized, 1);
4341 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4343 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4344 struct mem_cgroup_event *event, *tmp;
4347 * Unregister events and notify userspace.
4348 * Notify userspace about cgroup removing only after rmdir of cgroup
4349 * directory to avoid race between userspace and kernelspace.
4351 spin_lock(&memcg->event_list_lock);
4352 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4353 list_del_init(&event->list);
4354 schedule_work(&event->remove);
4356 spin_unlock(&memcg->event_list_lock);
4358 vmpressure_cleanup(&memcg->vmpressure);
4360 memcg_deactivate_kmem(memcg);
4362 wb_memcg_offline(memcg);
4365 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4367 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4369 memcg_destroy_kmem(memcg);
4370 __mem_cgroup_free(memcg);
4374 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4375 * @css: the target css
4377 * Reset the states of the mem_cgroup associated with @css. This is
4378 * invoked when the userland requests disabling on the default hierarchy
4379 * but the memcg is pinned through dependency. The memcg should stop
4380 * applying policies and should revert to the vanilla state as it may be
4381 * made visible again.
4383 * The current implementation only resets the essential configurations.
4384 * This needs to be expanded to cover all the visible parts.
4386 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4388 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4390 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4391 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4392 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4394 memcg->high = PAGE_COUNTER_MAX;
4395 memcg->soft_limit = PAGE_COUNTER_MAX;
4396 memcg_wb_domain_size_changed(memcg);
4400 /* Handlers for move charge at task migration. */
4401 static int mem_cgroup_do_precharge(unsigned long count)
4405 /* Try a single bulk charge without reclaim first */
4406 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4408 mc.precharge += count;
4411 if (ret == -EINTR) {
4412 cancel_charge(root_mem_cgroup, count);
4416 /* Try charges one by one with reclaim */
4418 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4420 * In case of failure, any residual charges against
4421 * mc.to will be dropped by mem_cgroup_clear_mc()
4422 * later on. However, cancel any charges that are
4423 * bypassed to root right away or they'll be lost.
4426 cancel_charge(root_mem_cgroup, 1);
4436 * get_mctgt_type - get target type of moving charge
4437 * @vma: the vma the pte to be checked belongs
4438 * @addr: the address corresponding to the pte to be checked
4439 * @ptent: the pte to be checked
4440 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4443 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4444 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4445 * move charge. if @target is not NULL, the page is stored in target->page
4446 * with extra refcnt got(Callers should handle it).
4447 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4448 * target for charge migration. if @target is not NULL, the entry is stored
4451 * Called with pte lock held.
4458 enum mc_target_type {
4464 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4465 unsigned long addr, pte_t ptent)
4467 struct page *page = vm_normal_page(vma, addr, ptent);
4469 if (!page || !page_mapped(page))
4471 if (PageAnon(page)) {
4472 if (!(mc.flags & MOVE_ANON))
4475 if (!(mc.flags & MOVE_FILE))
4478 if (!get_page_unless_zero(page))
4485 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4486 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4488 struct page *page = NULL;
4489 swp_entry_t ent = pte_to_swp_entry(ptent);
4491 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4494 * Because lookup_swap_cache() updates some statistics counter,
4495 * we call find_get_page() with swapper_space directly.
4497 page = find_get_page(swap_address_space(ent), ent.val);
4498 if (do_swap_account)
4499 entry->val = ent.val;
4504 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4505 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4511 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4512 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4514 struct page *page = NULL;
4515 struct address_space *mapping;
4518 if (!vma->vm_file) /* anonymous vma */
4520 if (!(mc.flags & MOVE_FILE))
4523 mapping = vma->vm_file->f_mapping;
4524 pgoff = linear_page_index(vma, addr);
4526 /* page is moved even if it's not RSS of this task(page-faulted). */
4528 /* shmem/tmpfs may report page out on swap: account for that too. */
4529 if (shmem_mapping(mapping)) {
4530 page = find_get_entry(mapping, pgoff);
4531 if (radix_tree_exceptional_entry(page)) {
4532 swp_entry_t swp = radix_to_swp_entry(page);
4533 if (do_swap_account)
4535 page = find_get_page(swap_address_space(swp), swp.val);
4538 page = find_get_page(mapping, pgoff);
4540 page = find_get_page(mapping, pgoff);
4546 * mem_cgroup_move_account - move account of the page
4548 * @nr_pages: number of regular pages (>1 for huge pages)
4549 * @from: mem_cgroup which the page is moved from.
4550 * @to: mem_cgroup which the page is moved to. @from != @to.
4552 * The caller must confirm following.
4553 * - page is not on LRU (isolate_page() is useful.)
4554 * - compound_lock is held when nr_pages > 1
4556 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4559 static int mem_cgroup_move_account(struct page *page,
4560 unsigned int nr_pages,
4561 struct mem_cgroup *from,
4562 struct mem_cgroup *to)
4564 unsigned long flags;
4568 VM_BUG_ON(from == to);
4569 VM_BUG_ON_PAGE(PageLRU(page), page);
4571 * The page is isolated from LRU. So, collapse function
4572 * will not handle this page. But page splitting can happen.
4573 * Do this check under compound_page_lock(). The caller should
4577 if (nr_pages > 1 && !PageTransHuge(page))
4581 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
4582 * of its source page while we change it: page migration takes
4583 * both pages off the LRU, but page cache replacement doesn't.
4585 if (!trylock_page(page))
4589 if (page->mem_cgroup != from)
4592 anon = PageAnon(page);
4594 spin_lock_irqsave(&from->move_lock, flags);
4596 if (!anon && page_mapped(page)) {
4597 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4599 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4604 * move_lock grabbed above and caller set from->moving_account, so
4605 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4606 * So mapping should be stable for dirty pages.
4608 if (!anon && PageDirty(page)) {
4609 struct address_space *mapping = page_mapping(page);
4611 if (mapping_cap_account_dirty(mapping)) {
4612 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4614 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4619 if (PageWriteback(page)) {
4620 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4622 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4627 * It is safe to change page->mem_cgroup here because the page
4628 * is referenced, charged, and isolated - we can't race with
4629 * uncharging, charging, migration, or LRU putback.
4632 /* caller should have done css_get */
4633 page->mem_cgroup = to;
4634 spin_unlock_irqrestore(&from->move_lock, flags);
4638 local_irq_disable();
4639 mem_cgroup_charge_statistics(to, page, nr_pages);
4640 memcg_check_events(to, page);
4641 mem_cgroup_charge_statistics(from, page, -nr_pages);
4642 memcg_check_events(from, page);
4650 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4651 unsigned long addr, pte_t ptent, union mc_target *target)
4653 struct page *page = NULL;
4654 enum mc_target_type ret = MC_TARGET_NONE;
4655 swp_entry_t ent = { .val = 0 };
4657 if (pte_present(ptent))
4658 page = mc_handle_present_pte(vma, addr, ptent);
4659 else if (is_swap_pte(ptent))
4660 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4661 else if (pte_none(ptent))
4662 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4664 if (!page && !ent.val)
4668 * Do only loose check w/o serialization.
4669 * mem_cgroup_move_account() checks the page is valid or
4670 * not under LRU exclusion.
4672 if (page->mem_cgroup == mc.from) {
4673 ret = MC_TARGET_PAGE;
4675 target->page = page;
4677 if (!ret || !target)
4680 /* There is a swap entry and a page doesn't exist or isn't charged */
4681 if (ent.val && !ret &&
4682 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4683 ret = MC_TARGET_SWAP;
4690 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4692 * We don't consider swapping or file mapped pages because THP does not
4693 * support them for now.
4694 * Caller should make sure that pmd_trans_huge(pmd) is true.
4696 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4697 unsigned long addr, pmd_t pmd, union mc_target *target)
4699 struct page *page = NULL;
4700 enum mc_target_type ret = MC_TARGET_NONE;
4702 page = pmd_page(pmd);
4703 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4704 if (!(mc.flags & MOVE_ANON))
4706 if (page->mem_cgroup == mc.from) {
4707 ret = MC_TARGET_PAGE;
4710 target->page = page;
4716 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4717 unsigned long addr, pmd_t pmd, union mc_target *target)
4719 return MC_TARGET_NONE;
4723 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4724 unsigned long addr, unsigned long end,
4725 struct mm_walk *walk)
4727 struct vm_area_struct *vma = walk->vma;
4731 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4732 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4733 mc.precharge += HPAGE_PMD_NR;
4738 if (pmd_trans_unstable(pmd))
4740 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4741 for (; addr != end; pte++, addr += PAGE_SIZE)
4742 if (get_mctgt_type(vma, addr, *pte, NULL))
4743 mc.precharge++; /* increment precharge temporarily */
4744 pte_unmap_unlock(pte - 1, ptl);
4750 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4752 unsigned long precharge;
4754 struct mm_walk mem_cgroup_count_precharge_walk = {
4755 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4758 down_read(&mm->mmap_sem);
4759 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4760 up_read(&mm->mmap_sem);
4762 precharge = mc.precharge;
4768 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4770 unsigned long precharge = mem_cgroup_count_precharge(mm);
4772 VM_BUG_ON(mc.moving_task);
4773 mc.moving_task = current;
4774 return mem_cgroup_do_precharge(precharge);
4777 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4778 static void __mem_cgroup_clear_mc(void)
4780 struct mem_cgroup *from = mc.from;
4781 struct mem_cgroup *to = mc.to;
4783 /* we must uncharge all the leftover precharges from mc.to */
4785 cancel_charge(mc.to, mc.precharge);
4789 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4790 * we must uncharge here.
4792 if (mc.moved_charge) {
4793 cancel_charge(mc.from, mc.moved_charge);
4794 mc.moved_charge = 0;
4796 /* we must fixup refcnts and charges */
4797 if (mc.moved_swap) {
4798 /* uncharge swap account from the old cgroup */
4799 if (!mem_cgroup_is_root(mc.from))
4800 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4803 * we charged both to->memory and to->memsw, so we
4804 * should uncharge to->memory.
4806 if (!mem_cgroup_is_root(mc.to))
4807 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4809 css_put_many(&mc.from->css, mc.moved_swap);
4811 /* we've already done css_get(mc.to) */
4814 memcg_oom_recover(from);
4815 memcg_oom_recover(to);
4816 wake_up_all(&mc.waitq);
4819 static void mem_cgroup_clear_mc(void)
4822 * we must clear moving_task before waking up waiters at the end of
4825 mc.moving_task = NULL;
4826 __mem_cgroup_clear_mc();
4827 spin_lock(&mc.lock);
4830 spin_unlock(&mc.lock);
4833 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
4834 struct cgroup_taskset *tset)
4836 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4837 struct mem_cgroup *from;
4838 struct task_struct *p;
4839 struct mm_struct *mm;
4840 unsigned long move_flags;
4844 * We are now commited to this value whatever it is. Changes in this
4845 * tunable will only affect upcoming migrations, not the current one.
4846 * So we need to save it, and keep it going.
4848 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4852 p = cgroup_taskset_first(tset);
4853 from = mem_cgroup_from_task(p);
4855 VM_BUG_ON(from == memcg);
4857 mm = get_task_mm(p);
4860 /* We move charges only when we move a owner of the mm */
4861 if (mm->owner == p) {
4864 VM_BUG_ON(mc.precharge);
4865 VM_BUG_ON(mc.moved_charge);
4866 VM_BUG_ON(mc.moved_swap);
4868 spin_lock(&mc.lock);
4871 mc.flags = move_flags;
4872 spin_unlock(&mc.lock);
4873 /* We set mc.moving_task later */
4875 ret = mem_cgroup_precharge_mc(mm);
4877 mem_cgroup_clear_mc();
4883 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
4884 struct cgroup_taskset *tset)
4887 mem_cgroup_clear_mc();
4890 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4891 unsigned long addr, unsigned long end,
4892 struct mm_walk *walk)
4895 struct vm_area_struct *vma = walk->vma;
4898 enum mc_target_type target_type;
4899 union mc_target target;
4903 * We don't take compound_lock() here but no race with splitting thp
4905 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
4906 * under splitting, which means there's no concurrent thp split,
4907 * - if another thread runs into split_huge_page() just after we
4908 * entered this if-block, the thread must wait for page table lock
4909 * to be unlocked in __split_huge_page_splitting(), where the main
4910 * part of thp split is not executed yet.
4912 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4913 if (mc.precharge < HPAGE_PMD_NR) {
4917 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4918 if (target_type == MC_TARGET_PAGE) {
4920 if (!isolate_lru_page(page)) {
4921 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
4923 mc.precharge -= HPAGE_PMD_NR;
4924 mc.moved_charge += HPAGE_PMD_NR;
4926 putback_lru_page(page);
4934 if (pmd_trans_unstable(pmd))
4937 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4938 for (; addr != end; addr += PAGE_SIZE) {
4939 pte_t ptent = *(pte++);
4945 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4946 case MC_TARGET_PAGE:
4948 if (isolate_lru_page(page))
4950 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
4952 /* we uncharge from mc.from later. */
4955 putback_lru_page(page);
4956 put: /* get_mctgt_type() gets the page */
4959 case MC_TARGET_SWAP:
4961 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4963 /* we fixup refcnts and charges later. */
4971 pte_unmap_unlock(pte - 1, ptl);
4976 * We have consumed all precharges we got in can_attach().
4977 * We try charge one by one, but don't do any additional
4978 * charges to mc.to if we have failed in charge once in attach()
4981 ret = mem_cgroup_do_precharge(1);
4989 static void mem_cgroup_move_charge(struct mm_struct *mm)
4991 struct mm_walk mem_cgroup_move_charge_walk = {
4992 .pmd_entry = mem_cgroup_move_charge_pte_range,
4996 lru_add_drain_all();
4998 * Signal mem_cgroup_begin_page_stat() to take the memcg's
4999 * move_lock while we're moving its pages to another memcg.
5000 * Then wait for already started RCU-only updates to finish.
5002 atomic_inc(&mc.from->moving_account);
5005 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5007 * Someone who are holding the mmap_sem might be waiting in
5008 * waitq. So we cancel all extra charges, wake up all waiters,
5009 * and retry. Because we cancel precharges, we might not be able
5010 * to move enough charges, but moving charge is a best-effort
5011 * feature anyway, so it wouldn't be a big problem.
5013 __mem_cgroup_clear_mc();
5018 * When we have consumed all precharges and failed in doing
5019 * additional charge, the page walk just aborts.
5021 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5022 up_read(&mm->mmap_sem);
5023 atomic_dec(&mc.from->moving_account);
5026 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5027 struct cgroup_taskset *tset)
5029 struct task_struct *p = cgroup_taskset_first(tset);
5030 struct mm_struct *mm = get_task_mm(p);
5034 mem_cgroup_move_charge(mm);
5038 mem_cgroup_clear_mc();
5040 #else /* !CONFIG_MMU */
5041 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5042 struct cgroup_taskset *tset)
5046 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5047 struct cgroup_taskset *tset)
5050 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5051 struct cgroup_taskset *tset)
5057 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5058 * to verify whether we're attached to the default hierarchy on each mount
5061 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5064 * use_hierarchy is forced on the default hierarchy. cgroup core
5065 * guarantees that @root doesn't have any children, so turning it
5066 * on for the root memcg is enough.
5068 if (cgroup_on_dfl(root_css->cgroup))
5069 root_mem_cgroup->use_hierarchy = true;
5071 root_mem_cgroup->use_hierarchy = false;
5074 static u64 memory_current_read(struct cgroup_subsys_state *css,
5077 return mem_cgroup_usage(mem_cgroup_from_css(css), false);
5080 static int memory_low_show(struct seq_file *m, void *v)
5082 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5083 unsigned long low = READ_ONCE(memcg->low);
5085 if (low == PAGE_COUNTER_MAX)
5086 seq_puts(m, "max\n");
5088 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5093 static ssize_t memory_low_write(struct kernfs_open_file *of,
5094 char *buf, size_t nbytes, loff_t off)
5096 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5100 buf = strstrip(buf);
5101 err = page_counter_memparse(buf, "max", &low);
5110 static int memory_high_show(struct seq_file *m, void *v)
5112 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5113 unsigned long high = READ_ONCE(memcg->high);
5115 if (high == PAGE_COUNTER_MAX)
5116 seq_puts(m, "max\n");
5118 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5123 static ssize_t memory_high_write(struct kernfs_open_file *of,
5124 char *buf, size_t nbytes, loff_t off)
5126 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5130 buf = strstrip(buf);
5131 err = page_counter_memparse(buf, "max", &high);
5137 memcg_wb_domain_size_changed(memcg);
5141 static int memory_max_show(struct seq_file *m, void *v)
5143 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5144 unsigned long max = READ_ONCE(memcg->memory.limit);
5146 if (max == PAGE_COUNTER_MAX)
5147 seq_puts(m, "max\n");
5149 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5154 static ssize_t memory_max_write(struct kernfs_open_file *of,
5155 char *buf, size_t nbytes, loff_t off)
5157 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5161 buf = strstrip(buf);
5162 err = page_counter_memparse(buf, "max", &max);
5166 err = mem_cgroup_resize_limit(memcg, max);
5170 memcg_wb_domain_size_changed(memcg);
5174 static int memory_events_show(struct seq_file *m, void *v)
5176 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5178 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5179 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5180 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5181 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5186 static struct cftype memory_files[] = {
5189 .read_u64 = memory_current_read,
5193 .flags = CFTYPE_NOT_ON_ROOT,
5194 .seq_show = memory_low_show,
5195 .write = memory_low_write,
5199 .flags = CFTYPE_NOT_ON_ROOT,
5200 .seq_show = memory_high_show,
5201 .write = memory_high_write,
5205 .flags = CFTYPE_NOT_ON_ROOT,
5206 .seq_show = memory_max_show,
5207 .write = memory_max_write,
5211 .flags = CFTYPE_NOT_ON_ROOT,
5212 .seq_show = memory_events_show,
5217 struct cgroup_subsys memory_cgrp_subsys = {
5218 .css_alloc = mem_cgroup_css_alloc,
5219 .css_online = mem_cgroup_css_online,
5220 .css_offline = mem_cgroup_css_offline,
5221 .css_free = mem_cgroup_css_free,
5222 .css_reset = mem_cgroup_css_reset,
5223 .can_attach = mem_cgroup_can_attach,
5224 .cancel_attach = mem_cgroup_cancel_attach,
5225 .attach = mem_cgroup_move_task,
5226 .bind = mem_cgroup_bind,
5227 .dfl_cftypes = memory_files,
5228 .legacy_cftypes = mem_cgroup_legacy_files,
5233 * mem_cgroup_low - check if memory consumption is below the normal range
5234 * @root: the highest ancestor to consider
5235 * @memcg: the memory cgroup to check
5237 * Returns %true if memory consumption of @memcg, and that of all
5238 * configurable ancestors up to @root, is below the normal range.
5240 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5242 if (mem_cgroup_disabled())
5246 * The toplevel group doesn't have a configurable range, so
5247 * it's never low when looked at directly, and it is not
5248 * considered an ancestor when assessing the hierarchy.
5251 if (memcg == root_mem_cgroup)
5254 if (page_counter_read(&memcg->memory) >= memcg->low)
5257 while (memcg != root) {
5258 memcg = parent_mem_cgroup(memcg);
5260 if (memcg == root_mem_cgroup)
5263 if (page_counter_read(&memcg->memory) >= memcg->low)
5270 * mem_cgroup_try_charge - try charging a page
5271 * @page: page to charge
5272 * @mm: mm context of the victim
5273 * @gfp_mask: reclaim mode
5274 * @memcgp: charged memcg return
5276 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5277 * pages according to @gfp_mask if necessary.
5279 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5280 * Otherwise, an error code is returned.
5282 * After page->mapping has been set up, the caller must finalize the
5283 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5284 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5286 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5287 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5289 struct mem_cgroup *memcg = NULL;
5290 unsigned int nr_pages = 1;
5293 if (mem_cgroup_disabled())
5296 if (PageSwapCache(page)) {
5298 * Every swap fault against a single page tries to charge the
5299 * page, bail as early as possible. shmem_unuse() encounters
5300 * already charged pages, too. The USED bit is protected by
5301 * the page lock, which serializes swap cache removal, which
5302 * in turn serializes uncharging.
5304 if (page->mem_cgroup)
5308 if (PageTransHuge(page)) {
5309 nr_pages <<= compound_order(page);
5310 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5313 if (do_swap_account && PageSwapCache(page))
5314 memcg = try_get_mem_cgroup_from_page(page);
5316 memcg = get_mem_cgroup_from_mm(mm);
5318 ret = try_charge(memcg, gfp_mask, nr_pages);
5320 css_put(&memcg->css);
5322 if (ret == -EINTR) {
5323 memcg = root_mem_cgroup;
5332 * mem_cgroup_commit_charge - commit a page charge
5333 * @page: page to charge
5334 * @memcg: memcg to charge the page to
5335 * @lrucare: page might be on LRU already
5337 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5338 * after page->mapping has been set up. This must happen atomically
5339 * as part of the page instantiation, i.e. under the page table lock
5340 * for anonymous pages, under the page lock for page and swap cache.
5342 * In addition, the page must not be on the LRU during the commit, to
5343 * prevent racing with task migration. If it might be, use @lrucare.
5345 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5347 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5350 unsigned int nr_pages = 1;
5352 VM_BUG_ON_PAGE(!page->mapping, page);
5353 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5355 if (mem_cgroup_disabled())
5358 * Swap faults will attempt to charge the same page multiple
5359 * times. But reuse_swap_page() might have removed the page
5360 * from swapcache already, so we can't check PageSwapCache().
5365 commit_charge(page, memcg, lrucare);
5367 if (PageTransHuge(page)) {
5368 nr_pages <<= compound_order(page);
5369 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5372 local_irq_disable();
5373 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5374 memcg_check_events(memcg, page);
5377 if (do_swap_account && PageSwapCache(page)) {
5378 swp_entry_t entry = { .val = page_private(page) };
5380 * The swap entry might not get freed for a long time,
5381 * let's not wait for it. The page already received a
5382 * memory+swap charge, drop the swap entry duplicate.
5384 mem_cgroup_uncharge_swap(entry);
5389 * mem_cgroup_cancel_charge - cancel a page charge
5390 * @page: page to charge
5391 * @memcg: memcg to charge the page to
5393 * Cancel a charge transaction started by mem_cgroup_try_charge().
5395 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5397 unsigned int nr_pages = 1;
5399 if (mem_cgroup_disabled())
5402 * Swap faults will attempt to charge the same page multiple
5403 * times. But reuse_swap_page() might have removed the page
5404 * from swapcache already, so we can't check PageSwapCache().
5409 if (PageTransHuge(page)) {
5410 nr_pages <<= compound_order(page);
5411 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5414 cancel_charge(memcg, nr_pages);
5417 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5418 unsigned long nr_anon, unsigned long nr_file,
5419 unsigned long nr_huge, struct page *dummy_page)
5421 unsigned long nr_pages = nr_anon + nr_file;
5422 unsigned long flags;
5424 if (!mem_cgroup_is_root(memcg)) {
5425 page_counter_uncharge(&memcg->memory, nr_pages);
5426 if (do_swap_account)
5427 page_counter_uncharge(&memcg->memsw, nr_pages);
5428 memcg_oom_recover(memcg);
5431 local_irq_save(flags);
5432 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5433 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5434 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5435 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5436 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5437 memcg_check_events(memcg, dummy_page);
5438 local_irq_restore(flags);
5440 if (!mem_cgroup_is_root(memcg))
5441 css_put_many(&memcg->css, nr_pages);
5444 static void uncharge_list(struct list_head *page_list)
5446 struct mem_cgroup *memcg = NULL;
5447 unsigned long nr_anon = 0;
5448 unsigned long nr_file = 0;
5449 unsigned long nr_huge = 0;
5450 unsigned long pgpgout = 0;
5451 struct list_head *next;
5454 next = page_list->next;
5456 unsigned int nr_pages = 1;
5458 page = list_entry(next, struct page, lru);
5459 next = page->lru.next;
5461 VM_BUG_ON_PAGE(PageLRU(page), page);
5462 VM_BUG_ON_PAGE(page_count(page), page);
5464 if (!page->mem_cgroup)
5468 * Nobody should be changing or seriously looking at
5469 * page->mem_cgroup at this point, we have fully
5470 * exclusive access to the page.
5473 if (memcg != page->mem_cgroup) {
5475 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5477 pgpgout = nr_anon = nr_file = nr_huge = 0;
5479 memcg = page->mem_cgroup;
5482 if (PageTransHuge(page)) {
5483 nr_pages <<= compound_order(page);
5484 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5485 nr_huge += nr_pages;
5489 nr_anon += nr_pages;
5491 nr_file += nr_pages;
5493 page->mem_cgroup = NULL;
5496 } while (next != page_list);
5499 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5504 * mem_cgroup_uncharge - uncharge a page
5505 * @page: page to uncharge
5507 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5508 * mem_cgroup_commit_charge().
5510 void mem_cgroup_uncharge(struct page *page)
5512 if (mem_cgroup_disabled())
5515 /* Don't touch page->lru of any random page, pre-check: */
5516 if (!page->mem_cgroup)
5519 INIT_LIST_HEAD(&page->lru);
5520 uncharge_list(&page->lru);
5524 * mem_cgroup_uncharge_list - uncharge a list of page
5525 * @page_list: list of pages to uncharge
5527 * Uncharge a list of pages previously charged with
5528 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5530 void mem_cgroup_uncharge_list(struct list_head *page_list)
5532 if (mem_cgroup_disabled())
5535 if (!list_empty(page_list))
5536 uncharge_list(page_list);
5540 * mem_cgroup_migrate - migrate a charge to another page
5541 * @oldpage: currently charged page
5542 * @newpage: page to transfer the charge to
5543 * @lrucare: either or both pages might be on the LRU already
5545 * Migrate the charge from @oldpage to @newpage.
5547 * Both pages must be locked, @newpage->mapping must be set up.
5549 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5552 struct mem_cgroup *memcg;
5555 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5556 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5557 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5558 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5559 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5560 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5563 if (mem_cgroup_disabled())
5566 /* Page cache replacement: new page already charged? */
5567 if (newpage->mem_cgroup)
5571 * Swapcache readahead pages can get migrated before being
5572 * charged, and migration from compaction can happen to an
5573 * uncharged page when the PFN walker finds a page that
5574 * reclaim just put back on the LRU but has not released yet.
5576 memcg = oldpage->mem_cgroup;
5581 lock_page_lru(oldpage, &isolated);
5583 oldpage->mem_cgroup = NULL;
5586 unlock_page_lru(oldpage, isolated);
5588 commit_charge(newpage, memcg, lrucare);
5592 * subsys_initcall() for memory controller.
5594 * Some parts like hotcpu_notifier() have to be initialized from this context
5595 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5596 * everything that doesn't depend on a specific mem_cgroup structure should
5597 * be initialized from here.
5599 static int __init mem_cgroup_init(void)
5603 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5605 for_each_possible_cpu(cpu)
5606 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5609 for_each_node(node) {
5610 struct mem_cgroup_tree_per_node *rtpn;
5613 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5614 node_online(node) ? node : NUMA_NO_NODE);
5616 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5617 struct mem_cgroup_tree_per_zone *rtpz;
5619 rtpz = &rtpn->rb_tree_per_zone[zone];
5620 rtpz->rb_root = RB_ROOT;
5621 spin_lock_init(&rtpz->lock);
5623 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5628 subsys_initcall(mem_cgroup_init);
5630 #ifdef CONFIG_MEMCG_SWAP
5632 * mem_cgroup_swapout - transfer a memsw charge to swap
5633 * @page: page whose memsw charge to transfer
5634 * @entry: swap entry to move the charge to
5636 * Transfer the memsw charge of @page to @entry.
5638 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5640 struct mem_cgroup *memcg;
5641 unsigned short oldid;
5643 VM_BUG_ON_PAGE(PageLRU(page), page);
5644 VM_BUG_ON_PAGE(page_count(page), page);
5646 if (!do_swap_account)
5649 memcg = page->mem_cgroup;
5651 /* Readahead page, never charged */
5655 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5656 VM_BUG_ON_PAGE(oldid, page);
5657 mem_cgroup_swap_statistics(memcg, true);
5659 page->mem_cgroup = NULL;
5661 if (!mem_cgroup_is_root(memcg))
5662 page_counter_uncharge(&memcg->memory, 1);
5665 * Interrupts should be disabled here because the caller holds the
5666 * mapping->tree_lock lock which is taken with interrupts-off. It is
5667 * important here to have the interrupts disabled because it is the
5668 * only synchronisation we have for udpating the per-CPU variables.
5670 VM_BUG_ON(!irqs_disabled());
5671 mem_cgroup_charge_statistics(memcg, page, -1);
5672 memcg_check_events(memcg, page);
5676 * mem_cgroup_uncharge_swap - uncharge a swap entry
5677 * @entry: swap entry to uncharge
5679 * Drop the memsw charge associated with @entry.
5681 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5683 struct mem_cgroup *memcg;
5686 if (!do_swap_account)
5689 id = swap_cgroup_record(entry, 0);
5691 memcg = mem_cgroup_from_id(id);
5693 if (!mem_cgroup_is_root(memcg))
5694 page_counter_uncharge(&memcg->memsw, 1);
5695 mem_cgroup_swap_statistics(memcg, false);
5696 css_put(&memcg->css);
5701 /* for remember boot option*/
5702 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5703 static int really_do_swap_account __initdata = 1;
5705 static int really_do_swap_account __initdata;
5708 static int __init enable_swap_account(char *s)
5710 if (!strcmp(s, "1"))
5711 really_do_swap_account = 1;
5712 else if (!strcmp(s, "0"))
5713 really_do_swap_account = 0;
5716 __setup("swapaccount=", enable_swap_account);
5718 static struct cftype memsw_cgroup_files[] = {
5720 .name = "memsw.usage_in_bytes",
5721 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5722 .read_u64 = mem_cgroup_read_u64,
5725 .name = "memsw.max_usage_in_bytes",
5726 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5727 .write = mem_cgroup_reset,
5728 .read_u64 = mem_cgroup_read_u64,
5731 .name = "memsw.limit_in_bytes",
5732 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5733 .write = mem_cgroup_write,
5734 .read_u64 = mem_cgroup_read_u64,
5737 .name = "memsw.failcnt",
5738 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5739 .write = mem_cgroup_reset,
5740 .read_u64 = mem_cgroup_read_u64,
5742 { }, /* terminate */
5745 static int __init mem_cgroup_swap_init(void)
5747 if (!mem_cgroup_disabled() && really_do_swap_account) {
5748 do_swap_account = 1;
5749 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5750 memsw_cgroup_files));
5754 subsys_initcall(mem_cgroup_swap_init);
5756 #endif /* CONFIG_MEMCG_SWAP */